Towards Clinically Oriented Feature Detection for Melanoma: A Deep Learning Approach

The alarming rise in skin cancer incidence and mortality rates worldwide underscores the urgent need for effective diagnostic and treatment strategies. Traditional manual inspection of skin lesions, while widely practiced, is subject to limitations such as inter-observer variability and difficulties in early detection. The rapid advancements in Artificial Intelligence have paved the way for the development of Computer-Automated Diagnostic Systems to address these challenges. This study proposes a multi-label classification pipeline that leverages pre-trained deep learning architectures to generate explainable detection of features based on the 7-point checklist, with the aim of assisting dermatologists in making informed decisions regarding skin lesion malignancy. Comprehensive experiments are conducted to assess the performance of the proposed pipeline across various dermoscopic structures, demonstrating that it performs equivalently to or statistically significantly outperforms benchmark methods trained on a publicly available dataset. The integration of visually interpretable explanations highlights the salient regions that influence the decision-making process of the model. This explainability integration is crucial for clinical implementation, enabling clinicians to assess the reliability of the model's decisions and make informed judgments.