Advancing Dermatological AI: A Unified Deep Learning Pipeline for Skin Lesion Analysis

Abstract

Skin cancer is one of the most common types of cancer worldwide, and early and accurate diagnosis is required to avert the advancement of malignant cases. The current dermatology diagnosis system depends on expert human evaluation of dermoscopic images, which produces limited accessibility and inconsistent results. This study developed a comprehensive deep learning framework that combines lesion segmentation with multiple dataset classification, severity assessment, and explainable AI to improve diagnostic performance and clinical confidence. The system uses Attention U-Net for lesion segmentation, which produces a Dice coefficient of 0.9661 and IoU score of 0.9361 on the ISIC 2018 dataset. The classification module used EfficientNet models trained on the ISIC 2018 (seven classes) and PAD-UFES-20 (six classes) datasets and produced 87.68% and 73.33% accuracy, respectively. The severity prediction module, which uses clinical metadata attributes, provides a reliable assessment of lesion urgency. The heatmaps generated by the Grad-CAM visualization demonstrate decision-critical regions in an interpretable manner. The integrated system surpasses the performance of single components, while segmentation-assisted classification provides better results than raw image analysis. AI capabilities are matched to clinical demands through this unified approach.

Ähnliche Arbeiten