Figure - available from: Frontiers in Medicine
This content is subject to copyright.
Grading examples of the visual map explanation techniques. For asymmetry, the visual explanation map was divided into 4 quadrants and shape and color distribution were analyzed. If all four quadrants showed the same color and format, there was no asymmetry (0); if 2 or 3 quadrants were similar, there was mild asymmetry (1); and if all four quadrants were different, there was severe asymmetry (2). For borders, the clinical border area was compared to the highlighted visual map. If the visual technique showed no highlight or ≤ 50% of the border area highlighted with cold colors, it was considered as no highlight (0). If ≤ 50% of the area was highlighted with warm colors or >50% with cold colors, it was considered partial border highlight (1); if >50% of the areas was highlighted with warm colors, it was considered total border highlight (2). Finally, if >50% of the lesion's limits could not be evaluated clinically, it was considered non-available (N/A). For color abnormality, dermatologists decided to compare the most significant color abnormalities in the dermatoscopy image as if they had a saliency map in their minds, comparing the imaginary heatmaps to the ones in the visual techniques. If the clinical color abnormalities presented an agreement area of ≤ 75% for warm colors, it was considered total agreement (0); if it was 25−75% for warm colors or >75% for cold colors, it was considered as partial agreement (1); if it was < 25% for warm colors or 25−75% for cold colors, it was considered total disagreement (2). For grading the highlight colors, we established blue/purple as cold colors and orange/red as warm colors.
Source publication
Introduction
The use of deep convolutional neural networks for analyzing skin lesion images has shown promising results. The identification of skin cancer by faster and less expensive means can lead to an early diagnosis, saving lives and avoiding treatment costs. However, to implement this technology in a clinical context, it is important for spec...
Similar publications
Malignant melanoma (MM) is the most aggressive skin cancer, requiring early diagnosis for better outcomes. While deep learning models have shown promise in dermatological image analysis, their performance is constrained by limited training data. Generative Adversarial Networks (GANs) offer a solution by generating synthetic images for data augmenta...
Melanoma, a type of skin cancer originating in the pigment-producing cells (melanocytes), poses a significant public health concern due to its aggressive nature and potential for metastasis if left untreated. Early detection is pivotal for effective intervention and improved patient outcomes. Traditional methods of diagnosing skin cancer often rely...
Skin diseases are listed among the most frequently encountered diseases. Skin diseases such as eczema, melanoma, and others necessitate early diagnosis to avoid further complications. This study aims to enhance the diagnosis of skin disease by utilizing advanced image processing techniques and an attention-based vision approach to support dermatolo...
Due to the high incidence and possibly fatal nature of skin cancer, early identification is crucial for enhancing patient results. This paper presents a unique deep learning network, EfficientNetB0 ViT, to accurately classify skin lesions. The proposed method encompasses the scalability and efficiency of EfficientNetB0 with the global pattern recog...
Background
Difficulty obtaining a dermatological consultation is an obstacle to the early diagnosis of melanoma. On the one hand, patients survival depends on the lesion thickness at the time of diagnosis. On the other hand, dermatologists treat many patients with benign lesions. Optimizing patient care pathways is a major concern. The aim of the p...
Citations
... The method provides interpretability, computational efficiency, and works without requiring architectural modifications or correct model classification [73]. However, Eigen-CAM has limitations in practical applications: The heatmaps can be ambiguous, highlighting broad regions, which impedes fault localization in tasks requiring detailed analysis [80]. Eigen-CAM's computation of a linear combination of activations from convolutional layers to identify features along the first principal component may not capture non-linear relationships in complex data, resulting in less informative explanations [81]. ...
Fault identification in transmission line insulators is essential to keep the power system running. Using deep learning-based models combined with interpretative techniques can be an alternative to improve power grid inspections and increase their reliability. Based on that consideration, this paper proposes an optimized ensemble of deep learning models (OEDL) based on weighted boxes fusion (WBF), called OEDL-WBF, to enhance the fault detection of power grid insulators. The proposed model is hypertuned considering a tree-structured Parzen estimator (TPE), and interpretative results are provided using the eigenvector-based class activation map (Eigen-CAM) algorithm. The Eigen-CAM had better results than Grad-CAM, Activation-CAM, MaxActivation-CAM, and WeightedActivation-CAM. The multi-criteria optimization of the structure by TPE ensures that the appropriate hyperparameters of the you only look once (YOLO) model are used for object detection. With a mean average precision (mAP)@[0.5] of 0.9841 and mAP@[0.5:0.95] of 0.9722 the proposed OEDL-WBF outperforms other deep learning-based structures, such as YOLOv8, YOLOv9, YOLOv10, YOLOv11, and YOLOv12 in a benchmarking. The Eigen-CAM further helps to interpret the outcomes of the model.
... XAI techniques reveal the decision-making mechanisms of CNNs, which are not consistently transparent. Alternative methods like GradCAM and EigenCAM improve the decision-making process of AI, hence improving the validation of its final results [18]. ...
Typical manual processes are time‐consuming, error‐prone, and subjective, especially for complex radiological diagnoses. Although current artificial intelligence models show promising results for identifying caries, they generally fail due to a lack of well‐pre‐processed images. This research work is two‐fold. Initially, we propose a novel layer division non‐zero elimination model to reduce Poisson noise and de‐blur the acquired images. In the second step, we propose a more accurate and intuitive method in segmenting and classifying caries of the teeth. We used a total of 17 840 radiographs, which are a mix of bitewing and periapical X‐rays, for classification with ResNet‐50 and segmentation with ResUNet. ResNet‐50 uses skip connections within the residual blocks to solve the gradient issue existing in cavity presence. ResUNet combines the encoder‐decoder structure of U‐Net with the residual block features of ResNet to improve the performance of segmentation on radiographs with cavities. Finally, the Stochastic Gradient Descent optimizer was employed during the training phase to ensure the possibility of convergence and improve accuracy. ResNet‐50 was proven to outperform earlier versions, like ResNet‐18 and ResNet‐34, in achieving a recognition accuracy of 87% in the classification challenge, which is a very reliable indicator of promising results. Similarly, ResUNet was proved to be better than existing state‐of‐the‐art models such as CariesNet, DeepLab v3, and U‐Net++ in terms of accuracy, even achieving the level of 98% accuracy in segmentation.
... To enhance the transparency of the models' decision-making process, Gradientweighted Class Activation Mapping (Grad-CAM) was employed to visualize the models. By utilizing gradient information from the last convolutional layer for weighted fusion, class activation maps were generated to underscore key regions of the image pertinent to classification targets (21). ...
Purpose
Acute Mesenteric Ischemia (AMI) is a critical condition marked by restricted blood flow to the intestine, which can lead to tissue necrosis and fatal outcomes. We aimed to develop a deep learning (DL) model based on CT angiography (CTA) imaging and clinical data to diagnose AMI.
Methods
A retrospective study was conducted on 228 patients suspected of AMI, divided into training and test sets. Clinical data (medical history and laboratory indicators) was included in a multivariate logistic regression analysis to identify the independent factors associated with AMI and establish a clinical factors model. The arterial and venous CTA images were utilized to construct DL model. A Fusion Model was constructed by integrating clinical factors into the DL model. The performance of the models was assessed using receiver operating characteristic (ROC) curves and decision curve analysis (DCA).
Results
Albumin and International Normalized Ratio (INR) were associated with AMI by univariate and multivariate logistic regression (P < 0.05). In the test set, the area under ROC curve (AUC) of the clinical factor model was 0.60 (sensitivity 0.47, specificity 0.86). The AUC of the DL model based on CTA images reached 0.90, which was significantly higher than the AUC values of the clinical factor model, as confirmed by the DeLong test (P < 0.05). The Fusion Model also showed exceptional performance in terms of AUC, accuracy, sensitivity, specificity, and precision, with values of 0.96, 0.94, 0.94, 0.95, and 0.98, respectively. DCA indicated that the Fusion Model provided a greater net benefit than those of models based solely on imaging and clinical information across the majority of the reasonable threshold probabilities.
Conclusion
The incorporation of CTA images and clinical information into the model markedly enhances the diagnostic accuracy and efficiency of AMI. This approach provides a reliable tool for the early diagnosis of AMI and the subsequent implementation of appropriate clinical intervention.
... Advancements in ML and DL applications for melanoma detection are particularly noteworthy for their capacity to refine predictive accuracy through continuous learning. Deep learning models, such as convolutional neural networks (CNNs), excel in image recognition tasks, identifying subtle features like asymmetry, border irregularities, and color variations that signal malignancy [7]. With exposure to diverse datasets, these models demonstrate improved diagnostic precision, enhancing their potential for real-time use during patient consultations. ...
... XAI elucidates the decision-making process of CNNs, which may not always be transparent. Methods such as GradCAM and EigenCAM aid in comprehending the rationale behind AI decisions, thereby enhancing the validation of its results 26 . ...
This study evaluates the effectiveness of an Artificial Intelligence (AI)-based smartphone application designed for decay detection on intraoral photographs, comparing its performance to that of junior dentists. Conducted at The Aga Khan University Hospital, Karachi, Pakistan, this study utilized a dataset comprising 7,465 intraoral images, including both primary and secondary dentitions. These images were meticulously annotated by two experienced dentists and further verified by senior dentists. A YOLOv5s model was trained on this dataset and integrated into a smartphone application, while a Detection Transformer was also fine-tuned for comparative purposes. Explainable AI techniques were employed to assess the AI’s decision-making processes. A sample of 70 photographs was used to directly compare the application’s performance with that of junior dentists. Results showed that the YOLOv5s-based smartphone application achieved a precision of 90.7%, sensitivity of 85.6%, and an F1 score of 88.0% in detecting dental decay. In contrast, junior dentists achieved 83.3% precision, 64.1% sensitivity, and an F1 score of 72.4%. The study concludes that the YOLOv5s algorithm effectively detects dental decay on intraoral photographs and performs comparably to junior dentists. This application holds potential for aiding in the evaluation of the caries index within populations, thus contributing to efforts aimed at reducing the disease burden at the community level.
... Traditionally, Eigen-CAM has been utilized to elucidate the operational mechanisms of mostly convolutional neural networks by visually highlighting what they capture in images. This has proven invaluable in fields such as medical imaging [60,61], security [62], and pedestrian identification [63], among others. In our study, Eigen-CAM revealed the critical hourly information within the temporal blocks as interpreted by the 1D convolutional network. ...
Passive acoustic monitoring (PAM) is an effective, non-intrusive method for studying ecosystems, but obtaining meaningful ecological information from its large number of audio files is challenging. In this study, we take advantage of the expected animal behavior at different times of the day (e.g., higher acoustic animal activity at dawn) and develop a novel approach to use these time-based patterns. We organize PAM data into 24-hour temporal blocks formed with sound features from a pretrained VGGish network. These features feed a 1D convolutional neural network with a class activation mapping technique that gives interpretability to its outcomes. As a result, these diel-cycle blocks offer more accurate and robust hour-by-hour information than using traditional ecological acoustic indices as features, effectively recognizing key ecosystem patterns.
