Precision Segmentation and Binary Masking of Skin Lesions in Automated Dermatological Diagnostics Using Detectron2

Skin cancer is a common but preventable type of cancer. Observing the rising trend of cases being reported every year, accurate timely diagnose can stop the spread of disease in the body. Manual diagnosis requires clinical expertise and causes delays whereas automatic diagnosis can bring accuracy and save time. This research introduces an enhanced approach, Detectron2, a state-of-the-art modular framework for deep learning-based segmentation tasks that enables precise instance level segmentation, therefore, distinguishing overlapping lesions and accurately identifying individual regions of interests. This customized approach offers the ability to handle complex and irregular lesion structures, robustness against noise elements like hairlines and ink projections without any pre-processing. This paper utilized manually annotated International Skin Imaging Collaboration (ISIC) archive and PH2 datasets for training and testing. Detectron2 establishes itself as a promising tool by achieving Intersection over Union (IoU) scores of 0.917 for the 2016 ISIC dataset, 0.869 for the 2017 ISIC dataset, 0.942 for the 2018 ISIC dataset, and 0.915 for the PH2 dataset.