Feature set reduction and analysis for skin disease classification modeling using SVM-KNN and neural network

• Main Author: ZAMBALES, JOSHUA
• Format: text
• Published: Archīum Ateneo 2017
• Subjects: Skin > Diseases > Diagnosis; Diagnostic imaging > Data processing; Diagnostic imaging > Digital techniques; Image processing > Digital techniques; Support vector machines; Computer algorithms.
• Online Access: https://archium.ateneo.edu/theses-dissertations/61; http://rizalls.lib.admu.edu.ph/#section=resource&resourceid=1503024655&currentIndex=0&view=fullDetailsDetailsTab
• Institution: Ateneo De Manila University

In this study, two different approaches in classification of skin lesion using image processing techniques were compared. The first one used a fusion of SVM-KNN classifier while the other used a neural network. For pre-processing, the samples were filtered to reduce noise and then segmented to isolate the lesion area. The segmented lesions were then subjected to feature extraction. To do this, the samples were translated into a data set corresponding to color and texture features. Furthermore, the classification algorithms were paired with two different feature sets with one having 86 features (feature set A) and the other having 4,182 features (feature set B). The results of both models are as follows: using feature set A, the neural network scored an average F-measure of 86.49% percent while the SVM-KNN classifier scored only 70.49%. With feature set B, the neural network scored an verage F-measure of 46.84% while the SVM-KNN scored 50.56%. Adding unknown samples (hemangioma, shingles, tinea corporis) into the selection, the accuracy dropped but the order in terms of accuracy among the models did not change. Using 86 features, the neural network scored an average F-measure of 63.37% while the SVM-KNN scored 56.49%. Using feature set B, the SVM-KNN classifier scored an average F-measure of 42.01% while the neural network scored the lowest average F-measure of 34.37%. The experimentations showed that the classification is more accurate using feature set A regardless of the classification algorithm used. The addition of the RGB Histogram features in feature set B only introduced more similarities in the dataset as shown by the lower scores of the models.