DERMA TRACE: VISUAL REASONING–DRIVEN DEEP LEARNING FOR SKIN LESION MALIGNANCY ANALYSIS

Abstract

Skin cancer is one of the most prevalent and life-threatening diseases in the world, with most deaths from skin cancer being due to melanoma. Accurate and early identification of skin cancer is crucial to enhance the prognosis and mortality of patients. Although the automated deep learning of skin lesion classification has made great strides, most of the existing models are regarded as black-box models and are hard to understand and accept in the clinic. To overcome this challenge, this paper introduces Derma Trace as a visual reasoning–driven deep learning framework for skin lesion malignancy analysis, which can accurately and explainable analyse skin lesions. We suggest a CNN-based framework to efficiently extract local features and a transformer-based attention mechanism to capture global features in dermoscopic images. There is also a visual reasoning component that is embedded to detect and label the features on the lesions that might be useful for diagnosis, such as asymmetry, border irregularity, color variation, and structural patterns. The deep trace learning allows Derma Trace to not only enhance prediction accuracy, but also give clear visual evidence that helps guide clinical decisions. The framework is tested with benchmark datasets from dermatology, such as ISIC 2019 and HAM10000, with various types of benign and malignant skin lesions. The experimental results show that Derma Trace results outperform the traditional CNN-based method in classification accuracy, precision, recall, and F1-score. Moreover, Attention maps and Visual explanations generated by Grad-CAM are used to offer a visual understanding to the clinicians of the output of the model, thus improving the trust and reliability of automated diagnostic systems.