Segmentation of skin lesions and their attributes in dermatoscopic images based on convolutional neural networks

Abstract

Segmentation of skin lesions in dermoscopic images is an important step in diagnosing skin cancer, and Convolutional Neural Networks (CNNs) have emerged as powerful tools to address this challenge. This work evaluated the effectiveness of two CNN models, TernausNet-16 and Mask R-CNN, in segmenting skin lesions and five of their attributes in the dermatoscopic images from the ISIC 2018 Challenge dataset. Jaccard Similarity Index (JSI) and Dice Similarity Coefficient (DSC) have been used as evaluation metrics. The results revealed that Mask R-CNN significantly outperformed TernausNet-16. The best model achieved 82.57% JSI and 84.76% DSC for lesion segmentation, and 42.86% JSI and 51.60% DSC for attribute segmentation when addressing attribute imbalance. Despite the longer training time, the results highlighted the potential of Mask R-CNN for improving the effectiveness of melanoma segmentation.