Medical Image Segmentation for Skin Lesion Detection via Topological Data Analysis

According to the WHO, two individuals die every hour from skin cancer and about 9500 people get skin cancer every day just in the United States. Various computer vision algorithms have been introduced for skin lesion detection, classification, and segmentation. This paper proposes a new segmentation-based algorithm in order to select target components using the persistence diagram of the input images. The results, in comparison with the existing seven different both clustering- and histogram-based segmentation methods using three metrics, show improved performance. Medical image segmentation is an essential task in computer-aided diagnosis. The main improvement of our method is to detect one lesion component by changing the persistent diagram threshold more cautiously. Sparse matrix implementation by Ripser packages effectively computes 2594 training images in less than an hour. The experimental results on the ISIC dataset suggest that using our framework can improve the accuracy up to 88.57% and achieve advanced performance in the segmentation of skin lesions.