MULTI-LABEL PROTOTYPE BASED INTERPRETABLE MACHINE LEARNING FOR MELANOMA DETECTION
Afra Hussaindeen, Shehana Iqbal, Thanuja D. Ambegoda
Abstract
Skin cancer is the most common of all cancers that exist in the world and melanoma is the deadliest among all the skin cancers. It is found that melanoma roughly kills a person every hour somewhere in the world. Considering the severity of the disease, significant effort goes into minimizing delays in the process of diagnosing melanoma. There are several approaches based on Machine Learning (ML) that can assist dermatologists in melanoma detection. However, many experts hesitate to trust ML systems due to their black-box nature, despite the accuracy of their performance. This highlights the need for applications that facilitate not only accurate classifications but also the ability to justify such decisions. In this work, we propose a prototype-based interpretable melanoma detector that uses the Seven Point Checklist, a well-known criterion used for the detection of melanoma. Prototypes provide the justification behind the decisions suggested by the ML model in a way of showing similar cases that are already known. In addition to identifying the dermoscopic features listed in the seven-point checklist, our work aims to provide reasoning that is similar to the ones used by the dermatologists in clinical practice for each decision made by the model. F1-Score has been used as the main performance metric in evaluating the model performance and that of the best performing class was 0.87. Furthermore, we show comparisons of our approach with Local Interpretable Model-Agnostic Explanations (LIME), a popular approach for interpretability for deep learning models.