Effect of Segmentation on Skin Lesion Classification Using Multi-Scale Convolution Neural Networks

Addressing skin lesions is a significant medical challenge due to their complex structures. Early detection and treatment are crucial, as some lesions can be life-threatening. Recent research focuses on automatic classification models that can increase diagnostic accuracy. However, the efficiency of extracting region of interest (ROI) before categorization to increase prediction accuracy remains a serious hurdle. The proposed methodology utilizes DeepLabv3+ for segmentation, followed by a multi-scale classification system that categorizes lesions into three distinct classes: seborrheic keratosis, melanoma and nevus. Leveraging the International Skin Imaging Collaboration (ISIC) 2017 dataset, the study demonstrates that the segmentation-based approach achieves a 4% improvement in accuracy compared to models trained on non-segmented images.