Statistical Association Rules and Relevance Feedback: Powerful Allies to Improve the Retrieval of Medical Images.
ABSTRACT This work aims at developing an efficient support to improve the precision of medical image retrieval by content, introducing an approach that combines techniques of statistical association rule mining and relevance feedback. Low level features of shape and texture are extracted from images. Statistical association rules are used to select the most relevant features to discriminate the images, reducing the size of the feature vectors and eliminating noisy features that influence negatively the query results, making the whole process more efficient. Additionally, our approach uses a new relevance feedback technique to overcome the semantic gap that exists between low level features and the high level user interpretation of images. Experiments show that the combination of statistical association rule mining and the relevance feedback technique proposed here improve the precision of the query results up to 100%
- SourceAvailable from: Javier Ramírez[Show abstract] [Hide abstract]
ABSTRACT: A fundamental challenge that remains unsolved in the neuroimage field is the small sample size problem. Feature selection and extraction, which are based on a limited training set, are likely to display poor generalization performance on new datasets. To address this challenge, a novel voxel selection method based on association rule (AR) mining is proposed for designing a computer aided diagnosis (CAD) system. The proposed method is tested as a tool for the early diagnosis of Alzheimer’s disease (AD). Discriminant brain areas are selected from a single photon emission computed tomography (SPECT) or positron emission tomography (PET) databases by means of an AR mining process. Simultaneously activated brain regions in control subjects that consist of the set of voxels defining the antecedents and consequents of the ARs are selected as input voxels for posterior dimensionality reduction. Feature extraction is defined by a subsequent reduction of the selected voxels using principal component analysis (PCA) or partial least squares (PLS) techniques while classification is performed by a support vector machine (SVM). The proposed method yields an accuracy up to 91.75% (with 89.29% sensitivity and 95.12% specificity) for SPECT and 90% (with 89.33% sensitivity and 90.67% specificity) for PET, thus improving recently developed methods for early diagnosis of AD.Expert Systems with Applications 10/2012; 39(14):11766-11774. · 1.85 Impact Factor
- [Show abstract] [Hide abstract]
ABSTRACT: This paper proposes the method called Feature Selection based on the Compactness Measure from Scatterplots (FSCoMS) to select the best features extracted from medical images aiming at improving the effectiveness of Content-Based Image Retrieval. This feature selection algorithm consists in a compactness analysis of scatterplots to find the most relevant features providing high separability abilities. A high relevance value of a scatterplot means better predictability among of classes based on two features. We take advantage of this information to generate a ranking for features usefulness. We compared our method to two well-known feature selection methods using three real medical datasets. All of them were compared regarding the dimensionality of the final feature vector and the retrieval effectiveness measured by the precision and recall graphs. The performed experiments show that our method not only obtained the highest retrieval performance but also achieved the smallest number of demanded features (dimensionality) than the other methods analyzed.Healthcare Informatics, Imaging and Systems Biology (HISB), 2012 IEEE Second International Conference on; 01/2012
- [Show abstract] [Hide abstract]
ABSTRACT: An innovative way of object shape representation using Density Histogram of Feature Points (DHFP) is introduced and used in this paper. We have named this method Enhanced Density Histogram of Feature Points (EDHFP). We use silhouette images where the image region ξ consists of only those pixels that correspond to points on the object and have a value one (1) indicating “on” pixels. We count the number of “on” pixels in rectangle boundaries around a centroid, in the event that there are no “on” pixels in a rectangle boundary then the value is zero. The similarity level indicator is introduced to form part of vector a representation of the object shape. This method of image representation shows improved retrieval rate when compared to Density Histogram of Feature Points (DHFP) representation method. This method is capable of grouping object shapes with high probability of being similar using the similarity level indicator before calculation of similarity distance. Analytic analysis is done to justify our method, experiments are conducted and we compared our results with the object shape representation by DHFP to prove EDHFP's robustness.01/2012;