Skin Lesion Segmentation in Clinical Images Using Nonnegative Matrix Factorization and Mutual Information

Early detection of cutaneous melanoma is crucial for improving patient survival rates. Skin lesion segmentation is a critical step towards computer-aided diagnosis that is effective in melanoma detection. This work proposes a novel patch-based model that leverages a combination of non-negative matrix factorization and mutual information for segmenting cutaneous melanomas in clinical images. The proposed approach tends to lead to accurate skin lesion segmentations, and it was comprehensively evaluated in terms of its effectiveness based on three benchmark datasets (DermIS, DermNet, and DermQuest) using statistical analysis, ablation studies, and cross-dataset testing. The proposed method tends to improve on the segmentation results provided by established deep learning baselines (FCNet, FCN, PSPNet, SegNet, U-Net) and by recent advancements specifically designed for skin lesion segmentation (DSNet, FrCNet, and LGPNet).