No full-text available
To read the full-text of this research,
you can request a copy directly from the authors.
A Methodological approach to the classification of dermoscopy images
Abstract
In this paper a methodological approach to the classification of pigmented skin lesions in dermoscopy images is presented. First, automatic border detection is performed to separate the lesion from the background skin. Shape features are then extracted from this border. For the extraction of color and texture related features, the image is divided into various clinically significant regions using the Euclidean distance transform. This feature data is fed into an optimization framework, which ranks the features using various feature selection algorithms and determines the optimal feature subset size according to the area under the ROC curve measure obtained from support vector machine classification. The issue of class imbalance is addressed using various sampling strategies, and the classifier generalization error is estimated using Monte Carlo cross validation. Experiments on a set of 564 images yielded a specificity of 92.34% and a sensitivity of 93.33%.
... Note that using visible light images captured from smartphones for automatic melanoma detection is quite new. Most previous works focused on dermoscopic images that are captured in the well-controlled clinical environments with specialized equipments [22]- [24] (including [25], which uses a smartphone for dermoscopic image analysis). Dermoscopic images include features below the skin surface, which cannot be captured with normal cameras equipped in smartphones (Fig. 1). ...
... Therefore, another measurement is needed to determine if there is a color variation all over the lesion or the color varies uniformly from the center to the border. Inspired by the clinical research [24], we use a feature called Color Triangle (CT) [28]. Specifically, given a lesion region (represented by a binary image), we first compute the center of mass of the region. ...
... 2) Lesion Border Feature (LBF) (16 Features): To describe the irregularity of the border, we compute shape features such as compactness, solidity, convexity, and variance of distances from border points to centroid of lesion [24]. Border irregu-larity is useful for melanoma detection: mole with irregular border has a higher probability of being melanoma [18]. ...
We investigate the design of an entire mobile imaging system for early detection of melanoma. Different from previous work, we focus on smartphone-captured visible light images. Our design addresses two major challenges. First, images acquired using a smartphone under loosely-controlled environmental conditions may be subject to various distortions, and this makes melanoma detection more difficult. Second, processing performed on a smartphone is subject to stringent computation and memory constraints. In our work, we propose a detection system that is optimized to run entirely on the resource-constrained smartphone. Our system intends to localize the skin lesion by combining a lightweight method for skin detection with a hierarchical segmentation approach using two fast segmentation methods. Moreover, we study an extensive set of image features and propose new numerical features to characterize a skin lesion. Furthermore, we propose an improved feature selection algorithm to determine a small set of discriminative features used by the final lightweight system. In addition, we study the human-computer interface (HCI) design to understand the usability and acceptance issues of the proposed system.
... Automatic skin lesion segmentation is an essential component in computer-aided diagnosis (CAD) of melanoma [4] [5]. However, this is a very challenging task due to significant variations in location, shape, size, color and texture across different patients. ...
... The appearance of a dermoscopic image with color manipulation. The left image in the first row shows the original image, and the rest are the adjusted images by normalizing the contrast of each channel to [5,95] percentile window, respectively. lesion segmentation much more challenging. ...
... In order to address this issue and to use color information more effectively, we added the three channels from HSV color space and the lightness channel (L) from CIELAB color space, which separate luma (image intensity) from chroma (color information) to allow independent process on these two types of information. Figure 1 shows how a dermoscopic image can be adjusted by normalizing the contrast of each channel to [5,95] percentile window. No other pre-processing was performed, so the input dimension to our CDNN model is 192 × 256 × 7. ...
Automatic skin lesion segmentation on dermoscopic images is an essential step in computer-aided diagnosis of melanoma. However, this task is challenging due to significant variations of lesion appearances across different patients. This challenge is further exacerbated when dealing with a large amount of image data. In this paper, we extended our previous work by developing a deeper network architecture with smaller kernels to enhance its discriminant capacity. In addition, we explicitly included color information from multiple color spaces to facilitate network training and thus to further improve the segmentation performance. We extensively evaluated our method on the ISBI 2017 skin lesion segmentation challenge. By training with the 2000 challenge training images, our method achieved an average Jaccard Index (JA) of 0.765 on the 600 challenge testing images, which ranked itself in the first place in the challenge
... Index Terms-Skin Lesions, Melanoma, Segmentation, Fully Convolutional Networks (FCNs) I. INTRODUCTION Melanoma (also known as malignant melanoma) has one of the most rapidly increasing incidences in the world and has considerable mortality rate if left untreated (Rigel et al. 1996). Early diagnosis is particularly important because melanoma can be cured with early excision (Celebi et al. 2007, Capdehourat et al. 2011. ...
... Skin lesion images such as dermoscopy are commonly acquired as a non-invasive imaging technique for the in-vivo evaluation of pigmented skin lesions and play an important role in early diagnosis (Celebi et al. 2007). ...
Automatic skin lesion segmentation methods based on fully convolutional networks (FCNs) are regarded as the state-of-the-art for accuracy. When there are, however, insufficient training data to cover all the variations in skin lesions, where lesions from different patients may have major differences in size/shape/texture, these methods failed to segment the lesions that have image characteristics, which are less common in the training datasets. FCN-based semi-automatic segmentation methods, which fuse user-inputs with high-level semantic image features derived from FCNs offer an ideal complement to overcome limitations of automatic segmentation methods. These semi-automatic methods rely on the automated state-of-the-art FCNs coupled with user-inputs for refinements, and therefore being able to tackle challenging skin lesions. However, there are a limited number of FCN-based semi-automatic segmentation methods and all these methods focused on early-fusion, where the first few convolutional layers are used to fuse image features and user-inputs and then derive fused image features for segmentation. For early-fusion based methods, because the user-input information can be lost after the first few convolutional layers, consequently, the user-input information will have limited guidance and constraint in segmenting the challenging skin lesions with inhomogeneous textures and fuzzy boundaries. Hence, in this work, we introduce a hyper-fusion network (HFN) to fuse the extracted user-inputs and image features over multiple stages. We separately extract complementary features which then allows for an iterative use of user-inputs along all the fusion stages to refine the segmentation. We evaluated our HFN on ISIC 2017, ISIC 2016 and PH2 datasets, and our results show that the HFN is more accurate and generalizable than the state-of-the-art methods.
... PSLs are evaluated using the ABCD rule; further assessment by a specialist is required for these lesions [110]. By computing asymmetry, the authors in [101] were able to estimate the lesion's major and minor axes [111]. The region of the lesion was split into 2 parts, employing an axis derived from the longest diagonal vector corresponding to the Euclidean distance of the affected area [112]. ...
... Another study involved Artificial Neural Networks (ANN) utilized as classification entities [138] and the features withdrawn were border and color. In another study, a skin lesion classification formulated on SVM is presented [111]. The features withdrawn include shape, texture, and color; attaining a SE of 93% and a SP of 92%. ...
Skin cancer has been recognized as one of the most lethal and complex types of cancer for over a decade. The diagnosis of skin cancer is of paramount importance, yet the process is intricate and challenging. The analysis and modeling of human skin pose significant difficulties due to its asymmetrical nature, the visibility of dense hair, and the presence of various substitute characteristics. The texture of the epidermis is notably different from that of normal skin, and these differences are often evident in cases of unhealthy skin. As a consequence, the development of an effective method for monitoring skin cancer has seen little progress. Moreover, the task of diagnosing skin cancer from dermoscopic images is particularly challenging. It is crucial to diagnose skin cancer at an early stage, despite the high cost associated with the procedure, as it is an expensive process. Unfortunately, the advancement of diagnostic techniques for skin cancer has been limited. To address this issue, there is a need for a more accurate and efficient method for identifying and categorizing skin cancer cases. This involves the evaluation of specific characteristics to distinguish between benign and malignant skin cancer occurrences. We present and evaluate several techniques for segmentation, categorized into three main types: thresholding, edge-based, and region-based. These techniques are applied to a dataset of 200 benign and melanoma lesions from the Hospital Pedro Hispano (PH2) collection. The evaluation is based on twelve distinct metrics, which are designed to measure various types of errors with particular clinical significance. Additionally, we assess the effectiveness of these techniques independently for three different types of lesions: melanocytic nevi, atypical nevi, and melanomas. The first technique is capable of classifying lesions into two categories: atypical nevi and melanoma, achieving the highest accuracy score of 90.00% with the Otsu (3-level) method. The second technique also classifies lesions into two categories: common nevi and melanoma, achieving a score of 90.80% with the Binarized Sauvola method.
... Techniques such as Support Vector Machines (SVM), which aim to find the optimal boundary between classes for data classification, k-nearest neighbor (k-NN), a non-parametric method that classifies data based on the closest neighbors in the feature space, and logistic regression, a statistical method used for binary classification by estimating the probability of a given input belonging to a particular class, were explored for diagnostic support. However, these methods did not achieve satisfactory detection performance due to high intra-class variation (where the lesions of the same classes can appear very different) and low inter-class variation (where the lesions of different classes can appear similar) [9][10][11]. ...
... The selection of ELU as the activation function is based on its ability to combine characteristics of GELU, Swish, and ReLU6, which all have non-linear and linear properties. ELU operates non-linearly for negative values and linearly for positive values, as expressed in Equation (11). ...
Skin cancer is one of the most easily developed cancers and is continuously seeing an increased incidence rate. In this study, we propose a novel ABC ensemble model for skin lesion classification by leveraging the ABCD rule, which is commonly used in dermatology to evaluate lesion features such as asymmetry, border, color, and diameter. Our model consists of five distinct blocks, two of which focus on learning general image characteristics, while the remaining three focus on specialized features related to the ABCD rule. The final classification results are achieved through a weighted soft voting approach, where the generalization blocks are assigned higher weights to optimize performance. Through 15 experiments using various model configurations, we show that the weighted ABC ensemble model outperforms the baseline models, achieving the best performance with an accuracy of 0.9326 and an F1-score of 0.9302. Additionally, Grad-CAM analysis is employed to assess how each block in the ensemble focuses on distinct lesion features, further enhancing the interpretability and reliability of the model. Our findings demonstrate that integrating general image features with specific lesion characteristics improves classification performance, and that adjusting the soft voting weights yields optimal results. This novel model offers a reliable tool for early skin lesion diagnosis.
... The conventional CAD procedure for melanoma diagnosis based on machine learning involves multiple steps. These include image preprocessing (e.g., artifact removal and color correction), lesion segmentation, extraction of handcrafted feature sets, and classification using dedicated classifiers based on the application context [7][8][9][10][11]. Initially, Celebi et al. [8] introduced a method that automatically detects borders, extracts shape, color, and texture features from relevant regions, selects optimized features, and addresses class imbalance specifically in dermoscopy images of pigmented skin lesions. ...
... These include image preprocessing (e.g., artifact removal and color correction), lesion segmentation, extraction of handcrafted feature sets, and classification using dedicated classifiers based on the application context [7][8][9][10][11]. Initially, Celebi et al. [8] introduced a method that automatically detects borders, extracts shape, color, and texture features from relevant regions, selects optimized features, and addresses class imbalance specifically in dermoscopy images of pigmented skin lesions. Based on this work, researchers have proposed ways to take advantage of the variety of hand-crafted features of melanoma. ...
The 7-point checklist (7PCL) is widely used in dermoscopy to identify malignant melanoma lesions needing urgent medical attention. It assigns point values to seven attributes: major attributes are worth two points each, and minor ones are worth one point each. A total score of three or higher prompts further evaluation, often including a biopsy. However, a significant limitation of current methods is the uniform weighting of attributes, which leads to imprecision and neglects their interconnections. Previous deep learning studies have treated the prediction of each attribute with the same importance as predicting melanoma, which fails to recognize the clinical significance of the attributes for melanoma. To address these limitations, we introduce a novel diagnostic method that integrates two innovative elements: a Clinical Knowledge-Based Topological Graph (CKTG) and a Gradient Diagnostic Strategy with Data-Driven Weighting Standards (GD-DDW). The CKTG integrates 7PCL attributes with diagnostic information, revealing both internal and external associations. By employing adaptive receptive domains and weighted edges, we establish connections among melanoma's relevant features. Concurrently, GD-DDW emulates dermatologists' diagnostic processes, who first observe the visual characteristics associated with melanoma and then make predictions. Our model uses two imaging modalities for the same lesion, ensuring comprehensive feature acquisition. Our method shows outstanding performance in predicting malignant melanoma and its features, achieving an average AUC value of 85%. This was validated on the EDRA dataset, the largest publicly available dataset for the 7-point checklist algorithm. Specifically, the integrated weighting system can provide clinicians with valuable data-driven benchmarks for their evaluations.
... A mole is considered to be of greater size when its diameter exceeds 6 millimeters, an amount comparable in size to that of a pencil eraser [6]. Dermoscopy utilizes a handheld instrument to magnify and illuminate skin lesions, providing a non-invasive method [7]. When diagnosing melanoma, dermoscopy is a more precise method than visual inspection [8]. ...
Background: Skin cancer classification is a challenging task due to the fine-grained diversity in the appearance of various diagnostic categories. Detecting skin cancer at an early stage is vital for enhancing patient outcomes, as the prognosis for this condition greatly improves when diagnosed early. Convolutional neural networks have been found to be more effective than dermatologists in classifying multiclass skin cancer. Problem: The identification of skin cancer is frequently impeded by the subjective analysis of dermoscopic images, resulting in misdiagnoses and delayed treatments. The objective of this study is to create a reliable and effective classification system using the efficientnetb4 model, which will aid in early detection and ultimately enhance patient outcomes. Objective: The main goal of this study is to create a highly efficient and accurate classification system for skin cancer using the efficientnetb4 model. The goal of this system is to improve the accuracy of diagnoses, minimize misdiagnoses, and enable early detection of skin lesions, leading to better patient outcomes and a more efficient diagnostic process in dermatology. Methods: The EfficientNetB4 model is trained on the HAM10000 dataset using transfer learning and fine-tuning techniques on rotated images, zoomed in and out, and even flipped over to make variations. Then, it adjusted the hyperparameters in the fine-tuning step to fine-tune its weights so that the model could fit the classification task for skin lesions more precisely. Results: The leading model, EfficientNetB4, achieved a Top-1 Accuracy of 89.22%, a Top-2 accuracy of 88.82%, and a top-3 accuracy of 88.62%. Precision, recall, and F1 scores are computed for each class. This model has demonstrated excellent performance in melanoma (MEL) and benign kurtosis-like lesions (BKL). Criteria considering high-class imbalance were used in the assessment of Efficient Net classifiers. Models with an intermediate level of complexity, such as EfficientNetB4, demonstrated the most optimal performance. Confusion matrices were also discovered to be useful in identifying skin cancer varieties with the greatest capacity for generalization. Conclusion: Overall, EfficientNetB4 demonstrated superior performance in classifying multi-class skin cancer. Further development would be oversampling or synthetic data generation for even more class-balancing techniques to improve performance over underrepresented classes. More medical data, including images and clinical data, will probably increase the overall diagnostic accuracy.
... Early research was primarily based on traditional methods that combined handcrafted feature extraction and machine learning [6]. For instance, Celebi et al. [7] proposed an image segmentation strategy based on Euclidean distance transformation, while Abbas et al. [8] constructed color-texture feature models in color space. Although these methods achieved some success on limited datasets, they were constrained by inherent limitations such as complex feature engineering and weak generalization capabilities. ...
Early diagnosis of skin cancer can significantly improve patient survival. Currently, skin lesion classification faces challenges such as lesion–background semantic entanglement, high intra-class variability, artifactual interference, and more, while existing classification models lack modeling of physicians’ diagnostic paradigms. To this end, we propose DermViT, a medically driven deep learning architecture that addresses the above issues through a medically-inspired modular design. DermViT consists of three main modules: (1) Dermoscopic Context Pyramid (DCP), which mimics the multi-scale observation process of pathological diagnosis to adapt to the high intraclass variability of lesions such as melanoma, then extract stable and consistent data at different scales; (2) Dermoscopic Hierarchical Attention (DHA), which can reduce computational complexity while realizing intelligent focusing on lesion areas through a coarse screening–fine inspection mechanism; (3). Dermoscopic Feature Gate (DFG), which simulates the observation–verification operation of doctors through a convolutional gating mechanism and effectively suppresses semantic leakage of artifact regions. Our experimental results show that DermViT significantly outperforms existing methods in terms of classification accuracy (86.12%, a 7.8% improvement over ViT-Base) and number of parameters (40% less than ViT-Base) on the ISIC2018 and ISIC2019 datasets. Our visualization results further validate DermViT’s ability to locate lesions under interference conditions. By introducing a modular design that mimics a physician’s observation mode, DermViT achieves more logical feature extraction and decision-making processes for medical diagnosis, providing an efficient and reliable solution for dermoscopic image analysis.
... Furthermore, the class with the lowest number of instances is usually the class of interest from the point of view of the learning task [5]. This problem is of great interest because it turns up in many real-world classification problems, such as remotesensing [6], pollution detection [7], risk management [8], fraud detection [9], and especially medical diagnosis [10][11][12][13]. ...
In Data mining and Knowledge Discovery hidden and valuable knowledge from the data sources is discovered. The traditional algorithms used for knowledge discovery are bottle necked due to wide range of data sources availability. Class imbalance is a one of the problem arises due to data source which provide unequal class i.e. examples of one class in a training data set vastly outnumber examples of the other class(es). Researchers have rigorously studied several techniques to alleviate the problem of class imbalance, including resampling algorithms, and feature selection approaches to this problem. In this paper, we present a new hybrid frame work dubbed as Wrapper based Intelligent Under Sampling (WIUS) for learning from skewed training data. These algorithms provide a simpler and faster alternative by using C4.5 and wrapper as base algorithm. We conduct experiments using ten UCI data sets from various application domains using five algorithms for comparison on five evaluation metrics. Experimental results show that our method has higher Area under the ROC Curve, F-measure, precision, TP rate and TN rate values than many existing class imbalance learning methods.
... These systems generally use traditional computer vision techniques to ex-tract various aspects, like shape, color, and texture, and feed them into a classifier. [10,11,12,13]. ...
... Ganster et al. used radiometric and shape variables in to create a K-Nearest Neighbor (KNN) classifier that produced 87% sensitivity and 92% specificity [2]. Shape, color, and texture were among the 437 variables that M. Emre Celebi retrieved in [3]. Eighteen of these features have been determined to be optimal for teaching a support vector machine (SVM) to categorize lesions. ...
The human skin, which is the biggest organ and the outermost layer of the body, has seven layers that serve to shield interior organs. Because of its broad role in the integumentary system, maintaining the health of the skin is essential. Skin problems present substantial classification challenges for medical professionals since they include a wide spectrum of diseases, including dermatoses. Consequently, they are depending more and more on machine learning (ML) technologies to help them predict and categorize these diseases. In the field of imaging, convolutional neural networks (CNNs) have demonstrated performance that is comparable to, and in some cases surpasses, human capabilities. Within this research, we propose a novel CNN architecture designed to classify two specific skin diseases: Eczema (symptoms on legs and hands) and Seborrheic Keratoses (symptoms on ears and skin). Additionally, we compare the performance of six ML algorithms to determine the most accurate model. We trained and tested our proposed technique on the Dermnet 2021 DATASET, which consists of 2,332 pictures and is publically available on Kaggle. Our findings show that the suggested CNN model, which achieves an accuracy of 91.1% and an F1-score of 92.3%, surpasses other cutting-edge techniques. With an F1-score of 79.12% and an accuracy of 78.41%, Linear Regression (LR) was the most successful ML model that was examined.
... Automated border detection is often the first step in the automated or semi-automated analysis of dermoscopy images [7]. ...
Background: Dermoscopy is one of the major imaging modalities used in the diagnosis of melanoma and other pigmented skin lesions. Due to the difficulty and subjectivity of human interpretation, computerized analysis of dermoscopy images has become an important research area. One of the most important steps in dermoscopy image analysis is the automated detection of lesion borders. Methods: In this article, we present a systematic overview of the recent border detection methods in the literature paying particular attention to computational issues and evaluation aspects. Conclusion: Common problems with the existing approaches include the acquisition, size, and diagnostic distribution of the test image set, the evaluation of the results, and the inadequate description of the employed methods. Border determination by dermatologists appears to depend upon higher-level knowledge, therefore it is likely that the incorporation of domain knowledge in automated methods will enable them to perform better, especially in sets of images with a variety of diagnoses.
... The variety of automated image analysis techniques discussed is broad, but mostly restricted within the space of classical computer vision approaches, typically using combinations of low-level visual feature representations (color, edge, and texture descriptors, quantification of melanin based on color, etc.), rule-based image processing or segmentation algorithms, and classical machine learning techniques, such as k-nearest neighbor (kNN) and support vector machines (SVM). Some publications have presented algorithms that include segmentation of the lesion [16][17][18][19][20][21]. A team from the Pedro Hispano Hospital of Portugal sought to evaluate the performance of several (e.g., SVM and kNN) machine learning classifiers based on color, edge, and texture descriptors [22,23]. ...
Melanoma is the deadliest form of skin cancer. While curable with early detection, only highly trained specialists are capable of accurately recognizing the disease. As expertise is in limited supply, automated systems capable of identifying disease could save lives, reduce unnecessary biopsies, and reduce costs. Toward this goal, we propose a system that combines recent developments in deep learning with established machine learning approaches, creating ensembles of methods that are capable of segmenting skin lesions, as well as analyzing the detected area and surrounding tissue for melanoma detection. The system is evaluated using the largest publicly available benchmark dataset of dermoscopic images, containing 900 training and 379 testing images. New state-of-the-art performance levels are demonstrated, leading to an improvement in the area under receiver operating characteristic curve of 7.5% (0.843 vs. 0.783), in average precision of 4% (0.649 vs. 0.624), and in specificity measured at the clinically relevant 95% sensitivity operating point 2.9 times higher than the previous state-of-the-art (36.8% specificity compared to 12.5%). Compared to the average of 8 expert dermatologists on a subset of 100 test images, the proposed system produces a higher accuracy (76% vs. 70.5%), and specificity (62% vs. 59%) evaluated at an equivalent sensitivity (82%).
... Automated border detection is often the first step in the automated or semi-automated analysis of dermoscopy images [7]. It is crucial for the image analysis for two main reasons. ...
Background: Dermoscopy is one of the major imaging modalities used in the diagnosis of melanoma and other pigmented skin lesions. Due to the difficulty and subjectivity of human interpretation, dermoscopy image analysis has become an important research area. One of the most important steps in dermoscopy image analysis is the automated detection of lesion borders. Although numerous methods have been developed for the detection of lesion borders, very few studies were comprehensive in the evaluation of their results. Methods: In this paper, we evaluate five recent border detection methods on a set of 90 dermoscopy images using three sets of dermatologist-drawn borders as the ground-truth. In contrast to previous work, we utilize an objective measure, the Normalized Probabilistic Rand Index, which takes into account the variations in the ground-truth images. Conclusion: The results demonstrate that the differences between four of the evaluated border detection methods are in fact smaller than those predicted by the commonly used XOR measure.
... Based on the segmentation results, hand-crafted features can be extracted for melanoma recognition. Celebi et al. extracted several features including color and texture from segmented lesion region for skin lesion classification [9]. Schaefer used automatic border detection approach [10] to segment lesion area and then ensemble the extracted features, i.e. shape, texture and color, for melanoma recognition [11]. ...
Skin lesion is a severe disease in world-wide extent. Early detection of melanoma in dermoscopy images significantly increases the survival rate. However, the accurate recognition of melanoma is extremely challenging due to the following reasons, e.g. low contrast between lesions and skin, visual similarity between melanoma and non-melanoma lesions, etc. Hence, reliable automatic detection of skin tumors is very useful to increase the accuracy and efficiency of pathologists. International Skin Imaging Collaboration (ISIC) is a challenge focusing on the automatic analysis of skin lesion. In this paper, we proposed two deep learning methods to address all the three tasks announced in ISIC 2017, i.e. lesion segmentation (task 1), lesion dermoscopic feature extraction (task 2) and lesion classification (task 3). A deep learning framework consisting of two fully-convolutional residual networks (FCRN) is proposed to simultaneously produce the segmentation result and the coarse classification result. A lesion index calculation unit (LICU) is developed to refine the coarse classification results by calculating the distance heat-map. A straight-forward CNN is proposed for the dermoscopic feature extraction task. To our best knowledges, we are not aware of any previous work proposed for this task. The proposed deep learning frameworks were evaluated on the ISIC 2017 testing set. Experimental results show the promising accuracies of our frameworks, i.e. 0.718 for task 1, 0.833 for task 2 and 0.823 for task 3 were achieved.
... Third, relative color is more natural from a perceptual point of view. Recent studies [15,23,24] have confirmed the usefulness of relative color features in skin lesion image analysis. ...
Dermoscopy is a non-invasive skin imaging technique, which permits visualization of features of pigmented melanocytic neoplasms that are not discernable by examination with the naked eye. One of the most important features for the diagnosis of melanoma in dermoscopy images is the blue-white veil (irregular, structureless areas of confluent blue pigmentation with an overlying white "ground-glass" film). In this article, we present a machine learning approach to the detection of blue-white veil and related structures in dermoscopy images. The method involves contextual pixel classification using a decision tree classifier. The percentage of blue-white areas detected in a lesion combined with a simple shape descriptor yielded a sensitivity of 69.35% and a specificity of 89.97% on a set of 545 dermoscopy images. The sensitivity rises to 78.20% for detection of blue veil in those cases where it is a primary feature for melanoma recognition.
... In [5], the authors used an approach involving texture for skin lesions classification. Gray level Cooccurrence matrix (GLCM) is used for the extraction of textural features from the lesion region. ...
Melanoma, considered to be the most rapidly increasing cancer, has had great emphasis placed on its diagnosis. This paper proposes a powerful edge detection framework Self-adaptive Canny Edge Detection using Reinforcement Learning (CRL-Edge) that integrates Canny edge detection with reinforcement learning. This approach adaptively fine-tunes the threshold parameters of the canny so as to enhance the edge continuity particularly for images with weak boundaries. The research also focusses on proposing a feature extraction method Dominant Texture Color Patterns (DTCP) that effectively helps in classifying malignant melanoma from dermoscopic images. This method is proposed based on the extraction of texture and color features that are dominant in a particular local region. The RGB color channel that consists of texture patterns with more intensity variations is said to be a dominant texture feature and a color channel that has maximum color intensity variations is a dominant color feature. The texture-color patterns are combined together to form a pattern that is assigned a unique texture-color value that describes the image features. The proposed feature of texture and color is analyzed in dermoscopic color images to classify lesions as benign or malignant, using CatBoost, a gradient boosting technique. The CatBoost is compared with other gradient boosting algorithms like Random Forest, XGBoost and Light GBM. The experiments were conducted on two different databases, the ISIC Archive and the PH2 database. The images were evaluated, on the basis of performance metrics such as sensitivity, specificity, accuracy, precision, F1-score and AUC. The experiment results show that CRL-Edge segmentation provides better segmentation accuracy and the DTCP descriptor using CatBoost classifier provides enhanced classification accuracy for classifying malignant lesion. The new method is compared with different state-of-art methods and has demonstrated the best performance.
... Image segmentation is a critical task in medical image analysis, providing anatomical structure information essential for disease diagnosis and treatment planning [6,35]. Known deep learning (DL) methods have achieved state-ofthe-art (SOTA) performance in many medical image segmentation tasks, including convolutional neural network (CNN)-based methods [13,20,33,39,69], Transformerbased methods [5,21], and hybrid approaches [12,14,43,66]. ...
Deep learning (DL) methods have shown remarkable successes in medical image segmentation, often using large amounts of annotated data for model training. However, acquiring a large number of diverse labeled 3D medical image datasets is highly difficult and expensive. Recently, mask propagation DL methods were developed to reduce the annotation burden on 3D medical images. For example, Sli2Vol~\cite{yeung2021sli2vol} proposed a self-supervised framework (SSF) to learn correspondences by matching neighboring slices via slice reconstruction in the training stage; the learned correspondences were then used to propagate a labeled slice to other slices in the test stage. But, these methods are still prone to error accumulation due to the inter-slice propagation of reconstruction errors. Also, they do not handle discontinuities well, which can occur between consecutive slices in 3D images, as they emphasize exploiting object continuity. To address these challenges, in this work, we propose a new SSF, called \proposed, {for segmenting any anatomical structures in 3D medical images using only a single annotated slice per training and testing volume.} Specifically, in the training stage, we first propagate an annotated 2D slice of a training volume to the other slices, generating pseudo-labels (PLs). Then, we develop a novel Object Estimation Guided Correspondence Flow Network to learn reliable correspondences between consecutive slices and corresponding PLs in a self-supervised manner. In the test stage, such correspondences are utilized to propagate a single annotated slice to the other slices of a test volume. We demonstrate the effectiveness of our method on various medical image segmentation tasks with different datasets, showing better generalizability across different organs, modalities, and modals. Code is available at \url{https://github.com/adlsn/Sli2Volplus}
... Furthermore, recently, smartphone applications analyzing images obtained by standard cameras have proven effectiveness in describing and diagnosing diseases [11,12]. Therefore, an application based on a Computer-Aided Diagnosis (CAD) system using optical images could serve as a suitable and cost-effective alternative for melanoma diagnosis [13]. ...
Solar exposure behavior has led to a significant increase in melanoma cancer cases in recent years. The mortality rates caused by this disease are the highest among dermatological cancers. This is due to the complexity of diagnosis, doctors use the Case Based Reasoning approach to leverage similarity with previously diagnosed cases. While the CBR approach can effectively address complex cases, its reliance on potentially irrelevant features can limit its accuracy and introduce diagnostic errors. In this study, we introduce a novel approach that integrates CBR within a possibility theory framework to assist experts in melanoma early detection. To address the issue of non-informative features, we introduce a possibilistic selection block within our approach. This block enables the CBR system to focus solely on relevant features, thereby enhancing its accuracy. In this approach, possibility theory is intended to address the problem of ambiguity and uncertainty affecting skin images. Then, the most relevant features are selected based on a possibilistic formalism and used as key features in the similarity measure within the CBR approach. Experimental validation on two sets of optical and dermoscopic lesion image datasets illustrates that our approach can be appropriate for lesion severity classification. It achieves a specificity of 100% and an accuracy of 95% on both databases, surpassing recent existing methods in melanoma diagnosis.
... To extract the colour features from the concentrations of the Chromophore, we rely on the statistical parameters most commonly used by the AI community [3,6,23], for each concentration C j (p). These include the mean denoted C j (p), the standard deviation denoted σ j , the Signal-to-Noise Ratio denoted SNR j and the Entropy denoted E j , which are defined respectively by the following four equations: ...
We propose in this paper a new approach for identifying skin diseases from RGB dermatological images. Based on Blind Source Separation and Artificial Intelligence (AI), this approach proceeds in two steps. We begin by estimating the concentrations of the three main skin Chromophore separately, adopting a new source separation technique that exploits both their spatial sparsity and their positivity. We then utilize these concentrations to extract the more relevant features for classification in our second step using AI, rather than those directly extracted from the three spectral bands of the image, as used by most existing methods. The results of tests performed on two different databases of RGB dermatological images of melanoma and nevus demonstrate the superiority of our approach, in terms of melanoma identification, compared to two types of existing methods based on AI.
... Murugan et al. (Murugan et al. 2019) employed Gaussian filters for feature extraction from skin lesions, subsequently using SVM for classification. Some studies explored algorithmic feature selection, combining feature sets to enhance classification accuracy (Celebi et al. 2007). Xie et al. (Xie et al. 2016) expanded their study to include features like color, texture, and border traits, using a neural network ensemble for classification and PCA for dimension reduction. ...
This research tackles the problem of early melanoma detection, a critical concern due to the aggressive nature of the disease often initiated by deoxyribonucleic acid (DNA) damage from ultraviolet radiation. Early detection is vital for increasing survival rates, yet existing models face challenges like unbalanced classification and sensitivity to initial weight settings. Our study introduces an integrated method that enhances detection capabilities using an ensemble of CNNs (convolutional neural networks) to analyze complex image features, combined with a novel mutual learning-based artificial bee colony (ML-ABC) algorithm for optimizing initial weight configurations. Furthermore, we employ a reinforcement learning approach to address data imbalance effectively. The ML-ABC algorithm uses a mutual learning strategy that adjusts potential food source positions based on the superior fitness of two selected individuals through a mutual learning factor. Our model conceptualizes the classification problem as a sequence of decision-making steps where each classification action is rewarded, with more rewards for the minority class to mitigate imbalance. Our approach was tested on the ISIC-2020 dataset and outperformed existing methods, achieving high accuracy (0.8661) and F-measure (0.8540). The proposed model represents a significant advancement in melanoma screening technologies, promising to elevate the success rates of early interventions.
... This work demonstrates increased performance at distinguishing benign lesions and melanoma images using the SMOTE oversampling approach. In various research [96,97], the SMOTE approach has been used to correct for the imbalance in the number of occurrences between classes by oversampling the retrieved features from the imbalanced dataset. In the present study, researchers utilize a specific technique to address the imbalanced representation of melanoma and nevus class data by selecting features based on the MI criterion. ...
Even if more and more high-quality public datasets are available, one of the biggest problems with deep learning for skin lesion diagnosis is the scarcity of training samples. Deep Convolutional Neural Networks (CNNs) have exhibited remarkable performance across multiple Computer Vision tasks. Despite their tendency to surpass conventional machine learning methods, the models produced by these algorithms require larger datasets to be accessible during training. An increasingly popular method to deal with this problem is data augmentation, especially in domains like medical imaging where big datasets are hard to come by. Experiments on independent, unknown datasets have demonstrated improved performance using data augmentation, which creates more data to train the model. This survey devotes a lot of time to using augmentation techniques based on basic data augmentation algorithms and the pros and cons of Test Time Augmentation (TTA), Generative Adversarial Networks (GANs), Variational Auto-Encoders (VAEs), and Synthetic Minority Oversampling Techniques (SMOTE). The main focus of this survey is data augmentation, which addresses the problem of limited data by providing a solution within the data space. When these approaches are incorporated into clinical practice, our research attempts to clarify them and increase trust in the reliability of AI models created from augmented data.
... The aim of lesion segmentation is to separate lesion regions from normal skin regions at the pixel level. Earlier conventional methods [1][2][3] performed complex preprocessing of images and used manually created features to learn lesion segmentation, but these features are not effective in segmenting large lesions that are highly variable. Deep learning based methods can perform the task of lesion segmentation efficiently and automatically compared to traditional methods. ...
Automatic skin lesion segmentation from dermoscopy images is of great significance in the early treatment of skin cancers, which is yet challenging even for dermatologists due to the inherent issues, i.e., considerable size, shape and color variation, and ambiguous boundaries. In this paper, we propose a network BSP-Net that implements the combination of critical boundary information and segmentation tasks to simultaneously solve the variation and boundary problems in skin lesion segmentation. The architecture of BSP-Net primarily consists of a multi-scale boundary enhancement (MBE) module and a progressive fusion decoder (PD). The MBE module, by deeply extracting boundary information in both multi-axis frequency and multi-scale spatial domains, generates precise boundary key-point prediction maps. This process not only accurately models local boundary information but also effectively retains global contextual information. On the other hand, the PD employs an asymmetric decoding strategy, guiding the generation of refined segmentation results by combining boundary-enhanced features rich in geometric details with global features containing semantic information about lesions. This strategy progressively fuses boundary and semantic information at different levels, effectively enabling high-performance collaboration between cross-level contextual features. To assess the effectiveness of BSP-Net, we conducted extensive experiments on two public datasets (ISIC-2016 &PH2, ISIC-2018) and one private dataset (XJUSKin). BSP-Net achieved Dice coefficients of 90.81, 92.41, and 83.88%, respectively. Additionally, it demonstrated precise boundary delineation with Average Symmetric Surface Distance (ASSD) scores of 7.96, 6.88, and 10.92%, highlighting its strong performance in skin lesion segmentation.
... Convolutional Neural Network (CNN) is the best and also extensively used algorithm in melanoma image analysis by many researchers [11]- [13]. Some researchers have focused on extracting color and texture features from segmented lesion region for melanoma classification [14], [15] and employing feature selection techniques [16]- [18]. Each technique aims to refine melanoma detection and minimize the false positive rate. ...
Cancer, characterized by the World Health Organization
(WHO), encompasses a wide range of diseases that
can manifest in any part of the body. Commonly known as
malignant tumors, cancer involves the rapid proliferation of
abnormal cells that exceed their usual boundaries. These cells
have the ability to infiltrate neighboring tissues and organs,
and can even metastasize to distant parts of the body, posing
a serious consequence to personal health. Our focus shifts to
melanoma, according to the WHO’s 2020 report, melanoma
ranks as the fifth most prevalent cancer globally, with 1.09
million new cases reported annually. Alarmingly, the incidence of
melanoma continues to rise each year. It is a type of skin cancer,
the main contributing factors such as exposure to ultraviolet
(UV) radiation and family history. The focus of our study is
to identify effective strategies, based on existing evidence, for
raising social awareness about melanoma, leveraging available
technologies and mobile applications. By analyzing recent data
on melanoma cases, mortality rates, and their distribution across
different races, age groups, and ethnicities, we plan to advocate
for early detection methods. This approach is an important in
prompting investigations into how advanced technologies can be
seamlessly integrated into personal healthcare to identify early
signs and symptoms. Our study intends to explore how advanced
technologies can be harnessed to facilitate early detection and
treatment. An integrated approach, involving heightened awareness of personal health practices, advancements in melanoma diagnosis technologies, and the presence of skilled healthcare professionals, holds the potential to significantly mitigate the impact of melanoma on both individuals and communities.
... Melanoma skin cancer has always been one of the bulks of dangerous cancer diseases in the world, accounting for 75% of the main causes of skin cancer deaths [1,2]. The primary recognition of melanoma skin cancer enhances the survival rate and also treatment feasibility [3]. ...
Deep neural networks have shown promising results in the classification of skin lesion images, particularly when they focus on the most significant regions of an image. However, the identification of melanoma continues to pose a significant challenge, primarily because of the substantial variability both within and between classes in images of moles. A new framework is proposed in this paper to develop an innovative method for effective melanoma diagnosis, presenting a new approach for Skin Lesion Classification (SLC) that combines a modified Inception ResNet v2 and Efficient Net-B4 in an ensemble. In the modified Inception ResNet v2, we incorporate a Soft-Attention (SA) mechanism. The goal is to enhance the network’s performance by using the SA mechanism to highlight informative features and suppress those that induce noise. Given the diversity of skin lesions in terms of type, texture, color, shape, and distribution, our proposed method extracts crucial features from the images by employing both the SA and merging the features of two layers of Inception ResNet v2.Furthermore, we have modified EfficientNet-B4 by adding two dense layers to improve its classification capabilities. After classifying the images with these two models, we employ soft voting for the final classification. We evaluated the performance of our proposed method on the ISIC-2017 and ISIC-2018 datasets, considering metrics such as Accuracy, Precision, Recall, and F-measures. Experimental results and a comparison with existing techniques demonstrate that our method achieves higher accuracy than current state-of-the-art methods.
Skin cancer is one of the most frequent and deadly form of cancer. Essentially, it is an abnormal growth of skin cells that primarily occurs after contaminated hands with the sun. These days, it also appears on skin surfaces that are not exposed to sunlight. Skin cancer is smoothly curable only if it is diagnosed in its initial days. There are some prominent types of skin cancer named as melanoma, squamous cell carcinoma, basal cell carcinoma, and many others. Many machine learning and deep learning methods have been developed to interpret medical images, specifically those of skin lesions, it is difficult and tiresome to analyze these to find cancer manually. Computer-aided diagnosis systems have two essential procedures: classification and segmentation of lesions. These two procedures improve the quality of features retrieved from medical images. An overview of some methods used to diagnose skin cancer is provided to identify the most efficient preprocessing, segmentation, feature extraction, and classification of medical images. Various research methods for specific skin cancer classification are also explored in this study. A further hurdle to creating an optimal diagnosis algorithm is the absence of a dataset on skin cancer. In order to assist researchers in developing useful algorithms that rapidly and accurately diagnose skin cancer, the study offers to provide a current overview of the proposed solutions to the issues in skin cancer detection. We gathered the results in tabular form after analyzing the efficiency of the most recent research based on a variety of factors, including techniques, and the performance of the applied datasets. We have discussed the current Deep Learning and Machine Learning techniques for detecting skin cancer along with their limitations. Along with outlining the various assessment metrics, we have also discussed the research gaps and challenges, such as imbalanced datasets, intra-class variance, inter-class similarity, etc., in skin cancer detection. The survey demonstrates its superiority over various other surveys currently in use.
Methods for an early non-invasive diagnosis of melanoma using computer vision systems are considered. The existing computer vision systems using neural networks for classifying dermoscopic images do not allow tracking which diagnosis features are used to assign images to a particular class, reducing physicians’ trust in the results. As an alternative, an image analysis algorithm is proposed with the ability to provide justifications for the decisions made at each processing stage. The implementation of this algorithm is based on the medical algorithm of modified globular pattern analysis. A significant sign of malignancy in a neoplasm is its asymmetry. This criterion is widely used by physicians when performing visual assessment of skin neoplasms. However, up to now, the issues of evaluating the symmetry of globular patterns in artificial intelligence systems have not been fully studied and described. A method for evaluating the symmetry of globular patterns in artificial intelligence systems for diagnosing skin neoplasms has been developed. A dataset of dermoscopic images was formed, containing 50 images of each type of neoplasms with symmetrically and asymmetrically arranged globular patterns. Methods for isolating the neoplasm area and globules are described. A classification system based on a set of 12 quantitative symmetry characteristics has been developed. The Random Forest algorithm was used to classify images based on symmetry features. In the conducted experiment, a classification accuracy of 85% was achieved. The presented results contribute to the development of computer vision methods in dermatology and demonstrate the possibility of using the proposed method in clinical decision-making support systems when performing a modified analysis of dermoscopic patterns to diagnose skin neoplasms.
Early detection of melanoma, the deadliest form of skin cancer, is critical for effective treatment. Detecting skin lesions accurately from dermoscopic images remains challenging, with the pigment network being a crucial indicator for melanoma detection. The accurate identification of pigment networks in dermoscopic images is difficult due to image noise and unwanted hair artifacts, which can obscure meaningful diagnostic features. To address these challenges, this thesis proposes novel image processing approaches for computer-aided pigment network detection. These methods aim to enhance image clarity by filtering out unwanted elements and converting images into binary formats, facilitating more precise analysis. Initially, preprocessing involves filtering dermoscopic images to remove noise and unwanted hair artifacts, enhancing image quality for more effective analysis. Subsequently, binary conversion and creation of binary masks further refine image clarity. The final step involves detecting pigment networks and calculating key features such as diameter and radius to assess melanoma presence. The proposed method achieves comparable accuracy and efficiency to existing algorithms. It demonstrates an average precision of 0.89 and average recall of 0.87, with an overall accuracy of 85.5%.
In this paper, multivalued data or multiple values variables are defined. They are typical when there is some intrinsic uncertainty in data production, as the result of imprecise measuring instruments, such as in image recognition, in human judgments and so on. \noindent So far, contributions in symbolic data analysis literature provide data preprocessing criteria allowing for the use of standard methods such as factorial analysis, clustering, discriminant analysis, tree-based methods. As an alternative, this paper introduces a methodology for supervised classification, the so-called Dynamic CLASSification TREE (D-CLASS TREE), dealing simultaneously with both standard and multivalued data as well. For that, an innovative partitioning criterion with a tree-growing algorithm will be defined. Main result is a dynamic tree structure characterized by the simultaneous presence of binary and ternary partitions. A real world case study will be considered to show the advantages of the proposed methodology and main issues of the interpretation of the final results. A comparative study with other approaches dealing with the same types of data will be also shown. D-CLASS TREE outperforms its competitors in terms of accuracy, which is a fundamental aspect for predictive learning.
Methods for early non-invasive diagnosis of melanoma using computer vision systems are considered. Existing computer vision systems using neural networks for classifying dermoscopic images do not allow tracking which diagnostic features are used to assign images to a particular class, reducing physicians' trust in the results. As an alternative, an image analysis algorithm is proposed with the ability to present justifications for decisions made at each processing stage. The implementation of this algorithm is based on the medical algorithm of modified globular pattern analysis. A significant sign of malignancy in a neoplasm is its asymmetry. This criterion is widely used by doctors in visual assessment of skin neoplasms. However, currently, the issues of evaluating the symmetry of globular patterns in artificial intelligence systems are not fully studied and described. A method for evaluating the symmetry of globular patterns in artificial intelligence systems for diagnosing skin neoplasms has been developed. A dataset of dermoscopic images was formed, containing 50 images each of neoplasms with symmetrically and asymmetrically arranged globular patterns. Methods for isolating the neoplasm area and globules are described. A classification system based on a set of 12 quantitative symmetry characteristics has been developed. The Random Forest algorithm was used to classify images based on symmetry features. In the conducted experiment, a classification accuracy of 85% was achieved. The presented results contribute to the development of computer vision methods in dermatology and demonstrate the possibility of using the proposed method in clinical decision support systems for modified analysis of dermoscopic patterns for diagnosing skin neoplasms.
Skin cancer represents a significant global public health concern, with over five million new cases diagnosed annually. If not diagnosed at an early stage, skin diseases have the potential to pose a significant threat to human life. In recent years, deep learning has increasingly been used in dermatological diagnosis. In this paper, a multiclassification model based on the Inception-v2 network and the focal loss function is proposed on the basis of deep learning, and the ISIC 2019 dataset is optimised using data augmentation and hair removal to achieve seven classifications of dermatological images and generate heat maps to visualise the predictions of the model. The results show that the model has an average accuracy of 89.04%, a precision of 87.37%, recall of 90.15%, and an F1-score of 88.76%, The accuracy rates of ResNext101, MobileNetv2, Vgg19, and ConvNet are 88.50%, 85.30%, 88.57%, and 86.90%, respectively. These results show that our proposed model performs better than the above models and performs well in classifying dermatological images, which has significant application value.
Background
Cutaneous melanoma is the most aggressive form of skin cancer, responsible for most skin cancer-related deaths. Recent advances in artificial intelligence, jointly with the availability of public dermoscopy image datasets, have allowed to assist dermatologists in melanoma identification. While image feature extraction holds potential for melanoma detection, it often leads to high-dimensional data. Furthermore, most image datasets present the class imbalance problem, where a few classes have numerous samples, whereas others are under-represented.
Methods
In this paper, we propose to combine ensemble feature selection (FS) methods and data augmentation with the conditional tabular generative adversarial networks (CTGAN) to enhance melanoma identification in imbalanced datasets. We employed dermoscopy images from two public datasets, PH2 and Derm7pt, which contain melanoma and not-melanoma lesions. To capture intrinsic information from skin lesions, we conduct two feature extraction (FE) approaches, including handcrafted and embedding features. For the former, color, geometric and first-, second-, and higher-order texture features were extracted, whereas for the latter, embeddings were obtained using ResNet-based models. To alleviate the high-dimensionality in the FE, ensemble FS with filter methods were used and evaluated. For data augmentation, we conducted a progressive analysis of the imbalance ratio (IR), related to the amount of synthetic samples created, and evaluated the impact on the predictive results. To gain interpretability on predictive models, we used SHAP, bootstrap resampling statistical tests and UMAP visualizations.
Results
The combination of ensemble FS, CTGAN, and linear models achieved the best predictive results, achieving AUCROC values of 87% (with support vector machine and IR=0.9) and 76% (with LASSO and IR=1.0) for the PH2 and Derm7pt, respectively. We also identified that melanoma lesions were mainly characterized by features related to color, while not-melanoma lesions were characterized by texture features.
Conclusions
Our results demonstrate the effectiveness of ensemble FS and synthetic data in the development of models that accurately identify melanoma. This research advances skin lesion analysis, contributing to both melanoma detection and the interpretation of main features for its identification.
The increasing incidence of and resulting deaths associated with malignant skin tumors are a public health problem that can be minimized if detection strategies are improved. Currently, diagnosis is heavily based on physicians’ judgment and experience, which can occasionally lead to the worsening of the lesion or needless biopsies. Several non-invasive imaging modalities, e.g., confocal scanning laser microscopy or multiphoton laser scanning microscopy, have been explored for skin cancer assessment, which have been aligned with different artificial intelligence (AI) strategies to assist in the diagnostic task, based on several image features, thus making the process more reliable and faster. This systematic review concerns the implementation of AI methods for skin tumor classification with different imaging modalities, following the PRISMA guidelines. In total, 206 records were retrieved and qualitatively analyzed. Diagnostic potential was found for several techniques, particularly for dermoscopy images, with strategies yielding classification results close to perfection. Learning approaches based on support vector machines and artificial neural networks seem to be preferred, with a recent focus on convolutional neural networks. Still, detailed descriptions of training/testing conditions are lacking in some reports, hampering reproduction. The use of AI methods in skin cancer diagnosis is an expanding field, with future work aiming to construct optimal learning approaches and strategies. Ultimately, early detection could be optimized, improving patient outcomes, even in areas where healthcare is scarce.
Skin cancer (SC) is an important medical condition that necessitates prompt identification to ensure timely treatment. Although visual evaluation by dermatologists is considered the most reliable method, its efficacy is subjective and laborious. Deep learning-based computer-aided diagnostic (CAD) platforms have become valuable tools for supporting dermatologists. Nevertheless, current CAD tools frequently depend on Convolutional Neural Networks (CNNs) with huge amounts of deep layers and hyperparameters, single CNN model methodologies, large feature space, and exclusively utilise spatial image information, which restricts their effectiveness. This study presents SCaLiNG, an innovative CAD tool specifically developed to address and surpass these constraints. SCaLiNG leverages a collection of three compact CNNs and Gabor Wavelets (GW) to acquire a comprehensive feature vector consisting of spatial–textural–frequency attributes. SCaLiNG gathers a wide range of image details by breaking down these photos into multiple directional sub-bands using GW, and then learning several CNNs using those sub-bands and the original picture. SCaLiNG also combines attributes taken from various CNNs trained with the actual images and subbands derived from GW. This fusion process correspondingly improves diagnostic accuracy due to the thorough representation of attributes. Furthermore, SCaLiNG applies a feature selection approach which further enhances the model’s performance by choosing the most distinguishing features. Experimental findings indicate that SCaLiNG maintains a classification accuracy of 0.9170 in categorising SC subcategories, surpassing conventional single-CNN models. The outstanding performance of SCaLiNG underlines its ability to aid dermatologists in swiftly and precisely recognising and classifying SC, thereby enhancing patient outcomes.
Rare objects are often of great interest and great value. Until recently, however, rarity has not received much attention in the context of data mining. Now, as increasingly complex real-world problems are addressed, rarity, and the related problem of imbalanced data, are taking center stage. This article discusses the role that rare classes and rare cases play in data mining. The problems that can result from these two forms of rarity are described in detail, as are methods for addressing these problems. These descriptions utilize examples from existing research. So that this article provides a good survey of the literature on rarity in data mining. This article also demonstrates that rare classes and rare cases are very similar phenomena---both forms of rarity are shown to cause similar problems during data mining and benefit from the same remediation methods.
In this paper, the effect of grey level quantization on the ability of co-occurrence probability statistics to classify natural textures is studied. Generally, as a function of increasing grey levels, many of the statistics demonstrate a decrease in classification ability while a few maintain constant classification accuracy. None of the individual statistics show increasing classification accuracy throughout all grey levels. Correlation analysis is used to rationalize a preferred subset of statistics. The preferred statistics set (contrast, correlation, and entropy) is demonstrated to be an improvement over using single statistics or using the entire set of statistics. If the feature space dimension only allows for a single statistic, one of contrast, dissimilarity, inverse difference normalized, or inverse difference moment normalized, is recommended. Testing that compares (using all orientations separately), the average of all orientations and look direction averaging, when determining the co-occurrence features, indicates that the look direction or all orientations is preferred. The Fisher linear discriminant method is used for all classification testing. The Fisher criterion is used as a separability index to provide insight into the classification results. Testing is performed on Brodatz imagery as well as two separate SAR sea-ice data sets.
This paper introduces concepts and algorithms of feature selection, surveys existing feature selection algorithms for classification and clustering, groups and compares different algorithms with a categorizing framework based on search strategies, evaluation criteria, and data mining tasks, reveals unattempted combinations, and provides guidelines in selecting feature selection algorithms. With the categorizing framework, we continue our efforts toward-building an integrated system for intelligent feature selection. A unifying platform is proposed as an intermediate step. An illustrative example is presented to show how existing feature selection algorithms can be integrated into a meta algorithm that can take advantage of individual algorithms. An added advantage of doing so is to help a user employ a suitable algorithm without knowing details of each algorithm. Some real-world applications are included to demonstrate the use of feature selection in data mining. We conclude this work by identifying trends and challenges of feature selection research and development.
As a result of advances in skin imaging technology and the development of suitable image processing techniques during the last decade, there has been a significant increase of interest in the computer-aided diagnosis of melanoma. Automated border detection is one of the most important steps in this procedure, since the accuracy of the subsequent steps crucially depends on it. In this paper, a fast and unsupervised approach to border detection in dermoscopy images of pigmented skin lesions based on the Statistical Region Merging algorithm is presented. The method is tested on a set of 90 dermoscopy images. The border detection error is quantified by a metric in which a set of dermatologist-determined borders is used as the ground-truth. The proposed method is compared to six state-of-the-art automated methods (optimized histogram thresholding, orientation-sensitive fuzzy c-means, gradient vector flow snakes, dermatologist-like tumor extraction algorithm, meanshift clustering, and the modified JSEG method) and borders determined by a second dermatologist. The results demonstrate that the presented method achieves both fast and accurate border detection in dermoscopy images.
The tutorial starts with an overview of the concepts of VC dimension and structural risk minimization. We then describe linear Support Vector Machines (SVMs) for separable and non-separable data, working through a non-trivial example in detail. We describe a mechanical analogy, and discuss when SVM solutions are unique and when they are global. We describe how support vector training can be practically implemented, and discuss in detail the kernel mapping technique which is used to construct SVM solutions which are nonlinear in the data. We show how Support Vector machines can have very large (even infinite) VC dimension by computing the VC dimension for homogeneous polynomial and Gaussian radial basis function kernels. While very high VC dimension would normally bode ill for generalization performance, and while at present there exists no theory which shows that good generalization performance is guaranteed for SVMs, there are several arguments which support the observed high accuracy of SVMs, which we review. Results of some experiments which were inspired by these arguments are also presented. We give numerous examples and proofs of most of the key theorems. There is new material, and I hope that the reader will find that even old material is cast in a fresh light.
Several works point out class imbalance as an obstacle on applying machine learning algorithms to real world domains. However, in some cases, learning algorithms perform well on several imbalanced domains. Thus, it does not seem fair to directly correlate class imbalance to the loss of performance of learning algorithms. In this work, we develop a systematic study aiming to question whether class imbalances are truly to blame for the loss of performance of learning systems or whether the class imbalances are not a problem by themselves. Our experiments suggest that the problem is not directly caused by class imbalances, but is also related to the degree of overlapping among the classes.
An approach to the construction of classifiers from imbalanced datasets is
described. A dataset is imbalanced if the classification categories are not
approximately equally represented. Often real-world data sets are predominately
composed of "normal" examples with only a small percentage of "abnormal" or
"interesting" examples. It is also the case that the cost of misclassifying an
abnormal (interesting) example as a normal example is often much higher than
the cost of the reverse error. Under-sampling of the majority (normal) class
has been proposed as a good means of increasing the sensitivity of a classifier
to the minority class. This paper shows that a combination of our method of
over-sampling the minority (abnormal) class and under-sampling the majority
(normal) class can achieve better classifier performance (in ROC space) than
only under-sampling the majority class. This paper also shows that a
combination of our method of over-sampling the minority class and
under-sampling the majority class can achieve better classifier performance (in
ROC space) than varying the loss ratios in Ripper or class priors in Naive
Bayes. Our method of over-sampling the minority class involves creating
synthetic minority class examples. Experiments are performed using C4.5, Ripper
and a Naive Bayes classifier. The method is evaluated using the area under the
Receiver Operating Characteristic curve (AUC) and the ROC convex hull strategy.
For large, real-world inductive learning problems, the number of training
examples often must be limited due to the costs associated with procuring,
preparing, and storing the training examples and/or the computational costs
associated with learning from them. In such circumstances, one question of
practical importance is: if only n training examples can be selected, in what
proportion should the classes be represented? In this article we help to answer
this question by analyzing, for a fixed training-set size, the relationship
between the class distribution of the training data and the performance of
classification trees induced from these data. We study twenty-six data sets
and, for each, determine the best class distribution for learning. The
naturally occurring class distribution is shown to generally perform well when
classifier performance is evaluated using undifferentiated error rate (0/1
loss). However, when the area under the ROC curve is used to evaluate
classifier performance, a balanced distribution is shown to perform well. Since
neither of these choices for class distribution always generates the
best-performing classifier, we introduce a budget-sensitive progressive
sampling algorithm for selecting training examples based on the class
associated with each example. An empirical analysis of this algorithm shows
that the class distribution of the resulting training set yields classifiers
with good (nearly-optimal) classification performance.
Rare objects are often of great interest and great value. Until recently, however, rarity has not received much attention in the context of data mining. Now, as increasingly complex real-world problems are addressed, rarity, and the related problem of imbalanced data, are taking center stage. This article discusses the role that rare classes and rare cases play in data mining. The problems that can result from these two forms of rarity are described in detail, as are methods for addressing these problems. These descriptions utilize examples from existing research. So that this article provides a good survey of the literature on rarity in data mining. This article also demonstrates that rare classes and rare cases are very similar phenomena---both forms of rarity are shown to cause similar problems during data mining and benefit from the same remediation methods.
A system for the computerized analysis of images obtained from ELM has been developed to enhance the early recognition of malignant melanoma. As an initial step, the binary mask of the skin lesion is determined by several basic segmentation algorithms together with a fusion strategy. A set of features containing shape and radiometric features as well as local and global parameters is calculated to describe the malignancy of a lesion. Significant features are then selected from this set by application of statistical feature subset selection methods. The final kNN classification delivers a sensitivity of 87% with a specificity of 92%.
The analysis of cutaneous melanoma images by two coupled computers (IBM 7350/4361) was carried out on twenty color slides. Each color slide was digitized with a spatial reduction of 25 X 25 microns. Classic technics of digital image analysis and new algorithms were used to improve the contrast on the full image or a portion of it, contrast a skin lesion with statistical information deduced from another lesion, evaluate the shape of the lesion, the roughness of the surface, and the transition region from the lesion to the normal skin, and analyze a lesion from the chromatic point of view. The theoretical reasons of interest are to have an objective method that is easy to standardize and reliably repeatable and to be able to analyze details not perceivable by the human eye. If the same technic are used in the evaluation of histologic characteristics of the lesions, a chance of making much more sophisticated clinicopathologic correlations will be available. The system needs to be improved at the technical level so that the response time of acquisition of the digitized images is shortened by the use of a digital television camera and the development of new computer programs to be run on a small computer. Evaluation of the system's sensitivity and specificity and an adequate clinical trial are needed.
Background
It has been over a decade since the first mention was made of the computer as a tool for assisting clinicians in diagnosing skin lesions. This review tabulates and summarises the major research papers, and comments on the state of the field after a decade of research.
Conclusions: We conclude that epiluminescent microscopy has become the image‐capture technique of choice in this field. However, the reporting of research to date has been less than exemplary, making “reinvention of the wheel” likely. It also appears that although the goal of a clinically useful diagnostic system is closer, the complexity and variation displayed by skin lesions, coupled with the ad hoc direction and reporting of research, may hinder the achievement of this goal for some time to come.
The tutorial starts with an overview of the concepts of VC dimension and structural risk minimization. We then describe linear Support Vector Machines (SVMs) for separable and non-separable data, working through a non-trivial example in detail. We describe a mechanical analogy, and discuss when SVM solutions are unique and when they are global. We describe how support vector training can be practically implemented, and discuss in detail the kernel mapping technique which is used to construct SVM solutions which are nonlinear in the data. We show how Support Vector machines can have very large (even infinite) VC dimension by computing the VC dimension for homogeneous polynomial and Gaussian radial basis function kernels. While very high VC dimension would normally bode ill for generalization performance, and while at present there exists no theory which shows that good generalization performance is guaranteed for SVMs, there are several arguments which support the observed high accuracy of SVMs, which we review. Results of some experiments which were inspired by these arguments are also presented. We give numerous examples and proofs of most of the key theorems. There is new material, and I hope that the reader will find that even old material is cast in a fresh light.
This paper introduces concepts and algorithms of feature selection, surveys existing feature selection algorithms for classification and clustering, groups and compares different algorithms with a categorizing framework based on search strategies, evaluation criteria, and data mining tasks, reveals unattempted combinations, and provides guidelines in selecting feature selection algorithms. With the categorizing framework, we continue our efforts toward building an integrated system for intelligent feature selection. A unifying platform is proposed as an intermediate step. An illustrative example is presented to show how existing feature selection algorithms can be integrated into a meta algorithm that can take advantage of individual algorithms. An added advantage of doing so is to help a user employ a suitable algorithm without knowing details of each algorithm. Some real-world applications are included to demonstrate the use of feature selection in data mining. We conclude this work by identifying trends and challenges of feature selection research and development.
Plots, probability plots and regression tests tests using moments other tests for univariate normality goodness of fit tests testing for outliers in univariate samples power comparisons for univariate tests for normality testing for normalitywith censored data assessing multivariate normality testing for multivariate outliers testing for normal mixtures robust methods computational methods and issues. Appendices: data sets used in examples critical values for tests.
In the context of machine learning from examples this paper deals with the problem of estimating the quality of attributes with and without dependencies between then. Greedy search prevents current inductive machine learning algorithms to detect significant dependencies between the attributes. Recently, Kira and R.endell developed the RELIEF algorithm for estimating the quality of attributes that is able to detect dependencies between attributes. We show strong relation between R.ELIEF’s estimates and impurity functions, that are usually used for heuristic guidance of inductive learning algorithms. We propose to use RELIEFF, an extended version of RELIEF, instead of myopic impurity functions. We have reimplemented Assistant, a system for top down induction of decision trees, using RELIEFF as an estimator of attributes at each selection step. The algorithm is tested on several artificial and several real world problems. Results show the advantage of the presented approach to inductive learning and open a wide range of possibilities for using RELIEFF.
As a result of the advances in skin imaging technology and the development of suitable image processing techniques, during the last decade, there has been a significant increase of interest in the computer-aided diagnosis of skin cancer. Dermoscopy is a non-invasive skin imaging technique which permits visualization of features of pigmented melanocytic neoplasms that are not discernable by examination with the naked eye. One of the useful features in dermoscopic diagnosis is the blue-white veil (irregular, structureless areas of confluent blue pigmentation with an overlying white "ground-glass" film) which is mostly associated with invasive melanoma. In this preliminary study, a machine learning approach to the detection of blue-white veil areas in dermoscopy images is presented. The method involves pixel classification based on relative and absolute color features using a decision tree classifier. Promising results were obtained on a set of 224 dermoscopy images.
Dermoscopy is a critical tool in the evaluation of pigmented lesions, but understanding the wide array of patterns can be confusing, especially to the novice. This book presents a concise, organized approach to dermoscopy. The early chapters orient the clinician to the use of the scopes and typical images produced with oil immersion. The second chapter outlines the distinct patterns, ie, radial streaming, blue white veil, and psuedopods. The authors discuss the relative specificity and sensitivity of these patterns in relation to the diagnosis of melanoma vs dysplastic nevi and pigmented basal cell carcinoma. Most important, they provide a visual comparison of the dermoscopy image, the clinical presentation, and the histopathology. The later chapters in the book are devoted to specific lesions including melanocytic nevi, melanoma, and nonmelanocytic pigmented tumors. Each subject is presented with an extensive array of dermoscopic images covering almost every pattern variation. The melanoma chapter is very comprehensive, and tables are used to outline the microscopic features of melanoma. There are also tables that summarize the critical points when evaluating basal cell carcinomas and seborreic keratoses vs melanocytic lesions.
Color image quantization is a process of representing an image with a small number of well selected colors. In this article an algorithm for multidimensional data clustering (termed the variance-based algorithm), based on the criterion of minimization of the sum-of-squared error, is applied to the problem of reducing the number of colors used to represent a given color image. The suitability of the sum-of-squared error criterion for measuring the similarity between the original and quantized images is examined using a digitized image and a computer-generated image. The experimental results indicate that this error measure is basically consistent with the perceived quality of the quantized image. The performance of the variance-based algorithm is compared with that of other algorithms for color image quantization in terms of quantization images generated using the colors selected by the variance-based and the mediancut algorithms are also presented.
Assuming white illumination and dichromatic reflectance, we propose new color models c
1
c
2
c
3 and l
1
l
2
l
3 invariant to the viewing direction, object geometry and shading. Further, it is shown that l
1
l
2
l
3 is also invariant to highlights. Further, a change in spectral power distribution of the illumination is considered to propose a new photometric color invariant m
1
m
2
m
3 for matte objects.
To evaluate photometric color invariant object recognition in practice, experiments have been carried out on a database consisting of 500 images taken from 3-D multicolored man-made objects.
On the basis of the reported theory and experimental results, it is shown that high object recognition accuracy is achieved by l
1
l
2
l
3 and hue H followed by c
1
c
2
c
3 and normalized colors rgb under the constraint of white illumination. Finally, it is shown that solely m
1
m
2
m
3 is invariant to a change in illumination color.
For real-world concept learning problems, feature selection is important to speed up learning and to improve concept quality. We review and analyze past approaches to feature selection and note their strengths and weaknesses. We then introduce and theoretically examine a new algorithm Rellef which selects relevant features using a statistical method. Relief does not depend on heuristics, is accurate even if features interact, and is noise-tolerant. It requires only linear time in the number of given features and the number of training instances, regardless of the target concept complexity. The algorithm also has certain limitations such as nonoptimal feature set size. Ways to overcome the limitations are suggested. We also report the test results of comparison between Relief and other feature selection algorithms. The empirical results support the theoretical analysis, suggesting a practical approach to feature selection for real-world problems.
Algorithms for feature selection fall into two broad categories: wrappers that use the learning algorithm itself to evaluate the usefulness of features and filters that evaluate features according to heuristics based on general characteristics of the data. For application to large databases, filters have proven to be more practical than wrappers because they are much faster. However, most existing filter algorithms only work with discrete classification problems. This paper describes a fast, correlation-based filter algorithm that can be applied to continuous and discrete problems. The algorithm often out-performs the well-known ReliefF attribute estimator when used as a preprocessing step for naive Bayes, instance-based learning, decision trees, locally weighted regression, and model trees. It performs more feature selection than ReliefF does-reducing the data dimensionality by fifty percent in most cases. Also, decision and model trees built from the preprocessed data are often significantly smaller.
DNA micro-arrays now permit scientists to screen thousands of genes simultaneously and determine whether those genes are active, hyperactive or silent in normal or cancerous tissue. Because these new micro-array devices generate bewildering amounts of raw data, new analytical methods must be developed to sort out whether cancer tissues have distinctive signatures of gene expression over normal tissues or other types of cancer tissues.
In this paper, we address the problem of selection of a small subset of genes from broad patterns of gene expression data, recorded on DNA micro-arrays. Using available training examples from cancer and normal patients, we build a classifier suitable for genetic diagnosis, as well as drug discovery. Previous attempts to address this problem select genes with correlation techniques. We propose a new method of gene selection utilizing Support Vector Machine methods based on Recursive Feature Elimination (RFE). We demonstrate experimentally that the genes selected by our techniques yield better classification performance and are biologically relevant to cancer.
In contrast with the baseline method, our method eliminates gene redundancy automatically and yields better and more compact gene subsets. In patients with leukemia our method discovered 2 genes that yield zero leave-one-out error, while 64 genes are necessary for the baseline method to get the best result (one leave-one-out error). In the colon cancer database, using only 4 genes our method is 98% accurate, while the baseline method is only 86% accurate.
The area under the ROC ( Receiver Operating Characteristics) curve, or simply AUC, has been traditionally used in medical diagnosis since the 1970s. It has recently been proposed as an alternative single-number measure for evaluating the predictive ability of learning algorithms. However, no formal arguments were given as to why AUC should be preferred over accuracy. In this paper, we establish formal criteria for comparing two different measures for learning algorithms and we show theoretically and empirically that AUC is a better measure ( defined precisely) than accuracy. We then reevaluate well-established claims in machine learning based on accuracy using AUC and obtain interesting and surprising new results. For example, it has been well-established and accepted that Naive Bayes and decision trees are very similar in predictive accuracy. We show, however, that Naive Bayes is significantly better than decision trees in AUC. The conclusions drawn in this paper may make a significant impact on machine learning and data mining applications.
Digital image analysis has been introduced into the diagnosis of skin lesions based on dermoscopic pictures.
To develop a computer algorithm for the diagnosis of melanocytic lesions and to compare its diagnostic accuracy with the results of established dermoscopic classification rules.
In the Department of Dermatology, University of Tuebingen, Germany, 837 melanocytic skin lesions were prospectively imaged by a dermoscopy video system in consecutive patients. Of these lesions, 269 were excised and examined by histopathology: 84 were classified as cutaneous melanomas and 185 as benign melanocytic naevi. The remaining 568 lesions were diagnosed by dermoscopy as benign. Digital image analysis was performed in all 837 benign and malignant melanocytic lesions using 64 different analytical parameters.
For lesions imaged completely (diameter < or = 12 mm), three analytical parameters were found to distinguish clearly between benign and malignant lesions, while in incompletely imaged lesions six parameters enabled differentiation. Based on the respective parameters and logistic regression analysis, a diagnostic computer algorithm for melanocytic lesions was developed. Its diagnostic accuracy was 82% for completely imaged and 84% for partially imaged lesions. All 837 melanocytic lesions were classified by established dermoscopic algorithms and the diagnostic accuracy was found to be in the same range (ABCD rule 78%, Menzies' score 83%, seven-point checklist 88%, and seven features for melanoma 81%).
A diagnostic algorithm for digital image analysis of melanocytic lesions can achieve the same range of diagnostic accuracy as the application of dermoscopic classification rules by experts. The present diagnostic algorithm, however, still requires a medical expert who is qualified to recognize cutaneous lesions as being of melanocytic origin.
Contenido: Caracterización continua de imagen; Caracterización digital de imagen; Procesamiento lineal discreto de dos dimenciones; Mejoramiento de imágenes; Análisis de imágenes; Software para precesamiento de imágenes.
As a result of advances in skin imaging technology and the development of suitable image processing techniques, during the last decade, there has been a significant increase of interest in the computer-aided diagnosis of melanoma. Automated border detection is one of the most important steps in this procedure, because the accuracy of the subsequent steps crucially depends on it.
In this article, we present a fast and unsupervised approach to border detection in dermoscopy images of pigmented skin lesions based on the statistical region merging algorithm.
The method is tested on a set of 90 dermoscopy images. The border detection error is quantified by a metric in which three sets of dermatologist-determined borders are used as the ground-truth. The proposed method is compared with four state-of-the-art automated methods (orientation-sensitive fuzzy c-means, dermatologist-like tumor extraction algorithm, meanshift clustering, and the modified JSEG method).
The results demonstrate that the method presented here achieves both fast and accurate border detection in dermoscopy images.
Epiluminescence microscopy (ELM) is a noninvasive technique that, by employing the optical phenomenon of oil immersion, makes subsurface structures of the skin accessible for in vivo examination and thus provides additional criteria for the clinical diagnosis of pigmented skin lesions. At present, almost all studies about the value and clinical importance of ELM are based on data derived from ELM experts (ie, dermatologists specifically trained in this technique). In the present study, we attempt to determine whether the clinical diagnosis of pigmented skin lesions is significantly improved using ELM and whether ELM-trained individuals and dermatologists not trained in this technique profit equally from this technique. Randomly selected histologically proven pigmented skin lesion specimens, photographed with (ELM) and without oil immersion (surface microscopy) were presented by slide projection to six ELM experts and 13 ELM nonexperts (ie, dermatologists not formally trained in ELM) for diagnosis. To evaluate the diagnostic performance of ELM experts and nonexperts with and without the oil immersion technique (ie, ELM vs surface microscopy), the following parameters were obtained: intraobserver and interobserver agreement by kappa statistics and sensitivity and specificity of diagnostic performance.
Our results show that by using the ELM technique the ELM experts reach a substantially better intraobserver agreement than nonexperts (median kappa, 0.56 vs 0.36). The interobserver agreement was markedly increased in the ELM experts group (average gain, 7%) but decreased in the ELM nonexperts group (average loss, 6%). The sensitivity of diagnosis was significantly increased in the ELM experts group (average gain, 10%), but decreased in the nonexperts group (average loss, 10%). Finally, the specificity of diagnosis was excellent in the ELM experts group, both with and without oil immersion (0.91) and was somewhat improved by ELM in the nonexperts group (0.77 vs 0.85).
We conclude that the ELM technique increases sensitivity in formally trained dermatologists, but may decrease the diagnostic ability in dermatologists not formally trained in the ELM technique. Consequently, formal broad-based training in ELM should be offered to the dermatologic community.
Epiluminescence microscopy (ELM) is a noninvasive technique by which the clinical diagnosis of pigmented skin lesions (PSL) can be improved. Many ELM criteria have been described, but their significance in the differential diagnosis of PSL has not yet been established.
The purpose of this study was to determine the value of ELM criteria in the differential diagnosis of PSL.
Two hundred one melanocytic PSL (61 common nevi, 60 dysplastic nevi, and 80 melanomas) were investigated with ELM for the presence of certain ELM criteria; their significance was determined by calculating the odds ratios.
Individual ELM criteria have different weights of significance in the differential diagnosis of melanocytic PSL. Selected patterns of ELM criteria adjusted to the distinct types of PSL considerably improve the diagnostic accuracy of melanocytic PSL.
The prevalence of certain distinct ELM criteria in a given melanocytic PSL has statistical value in differential diagnosis.
Recently, there has been a growing number of studies applying image processing techniques to analyze melanocytic lesions for atypia and possible malignancy and for total-body mole mapping. However, such lesions can be partially obscured by body hairs. None of these studies has fully addressed the problem of human hairs occluding the imaged lesions. In our previous study we designed an automatic segmentation program to differentiate skin lesions from the normal healthy skin, and learned that the program performed well with most of the images, the exception being those with hairs, especially dark thick hairs, covering part of the lesions. These thick dark hairs confused the program, resulting in unsatisfactory segmentation results. In this paper, we present a method to remove hairs from an image using a pre-processing program we have called DullRazor. This pre-processing step enables the segmentation program to achieve satisfactory results. DullRazor can be downloaded as shareware from http:/(/)www.derm.ubc.ca.
Differentiation of melanoma from melanocytic nevi is difficult even for skin cancer specialists. This motivates interest in computer-assisted analysis of lesion images.
Our purpose was to offer fully automatic differentiation of melanoma from dysplastic and other melanocytic nevi through multispectral digital dermoscopy.
At 4 clinical centers, images were taken of pigmented lesions suspected of being melanoma before biopsy. Ten gray-level (MelaFind) images of each lesion were acquired, each in a different portion of the visible and near-infrared spectrum. The images of 63 melanomas (33 invasive, 30 in situ) and 183 melanocytic nevi (of which 111 were dysplastic) were processed automatically through a computer expert system to separate melanomas from nevi. The expert system used either a linear or a nonlinear classifier. The "gold standard" for training and testing these classifiers was concordant diagnosis by two dermatopathologists.
On resubstitution, 100% sensitivity was achieved at 85% specificity with a 13-parameter linear classifier and 100%/73% with a 12-parameter nonlinear classifier. Under leave-one-out cross-validation, the linear classifier gave 100%/84% (sensitivity/specificity), whereas the nonlinear classifier gave 95%/68%. Infrared image features were significant, as were features based on wavelet analysis.
Automatic differentiation of invasive and in situ melanomas from melanocytic nevi is feasible, through multispectral digital dermoscopy.