Enhanced U-Net-Based Segmentation of Skin Lesions Using Multi-Year ISIC Datasets and Hybrid Training Pipeline

Abstract

The segmentation of medical images, particularly in the framework of skin lesion detection, is one of the key tools in the early diagnosis of melanoma. However, datasets with a single calendar year are often too homogenous to permit generalization, and the models will also favor overfitting. This paper proposes an enhanced U-Net architecture, trained on a combined set of the ISIC 2016, ISIC 2017, and ISIC 2018 corpora. By merging all of these datasets and using standardized preprocessing steps, such as resizing, normalizing, and augmentation, we enhance the diversity of data and strengthen the models. The resulting corpus consists of 5,494 pairs of images and masks, including 70 percent to be trained, 15 percent to be validated, and 15 percent to be tested. The model suggested several improvements, such as a compound DiceBCE loss, dropout regularization, and after-processing. Performance was measured over a range of quantitative measures, including the Dice coefficient, intersection over union, accuracy, specificity, and area under the receiver operating characteristic curve, with results showing Dice coefficients of 0.90 or above and excellent segmentation performance. These results support the claim that dataset integration across multiple years enhances the performance of a model and should be further adopted to integrate datasets across years in future studies to generate clinically valid artificial intelligence solutions.