Deep Learning-Based Dermoscopic Image Classification System for Robust Skin Lesion Analysis

Abstract

This paper introduces a sophisticated dermoscopic image classification system (DICS) leveraging deep learning techniques for accurate skin lesion classification. The DICS comprises four distinct modules: i) Skin Lesion Segmentation (SLS), ii) Feature Extraction (FE), iii) Feature Selection (FS), and iv) Image Classification (IC). The SLS module preprocesses the input dermoscopic image and employs a color k-means clustering approach for segmentation. Subsequently, in the FE module, three types of features are extracted, including 4th order Color Moments (CM), a statistical model based on Generalized Autoregressive Conditional Heteroscedasticity (GARCH), and texture features derived from Local Binary Patterns (LBP). The predominant features are then selected in the FS module using a statistical t-test. Finally, the IC module classifies dermoscopic images as normal or melanoma using a deep learning approach. The DICS demonstrates promising results, achieving 99% and 100% accuracy in normal/abnormal and benign/malignant classifications, respectively, when tested on the PH²database. This robust classification system has the potential to contribute significantly to the field of dermatological image analysis.