Detection and Removal of Assymmetrical Skin Lesions Using DU-Net for Patch Extraction

Abstract

This study presents DSeg-net, a novel method for accurately identifying and removing
pigmented skin lesions from dermoscopic images, crucial for timely diagnosis and management
of melanoma. DSeg-net combines deep convolutional neural networks, particularly YOLOv5, for
patch detection, asymmetrical patch contouring for edge preservation, and clustering techniques
for patch extraction. Additionally, it employs De Trop Noise Exclusion with in-painting to
eliminate hair from challenging dataset images. The method involves rigorous annotation of skin
images with lesions of varying sizes and shapes using rectangle bounding, followed by fine-
tuning YOLOv5 hyperparameters for high-confidence multiple lesion detection. Despite
complex textures and unclear boundaries, DSeg-net consistently detects and labels patches,
accurately segmenting areas of skin pathology. Evaluation on various datasets demonstrates that
the proposed segmentation techniques achieve an overall average accuracy of approximately
92% to 94%.