Exploring XAI Attention Maps to Investigate the Effect of Distance Metric and Lesion-Shaped Border Expansion Size for Effective Content-Based Dermatological Lesion Retrieval

Since automating the early diagnosis for skin cancer is becoming a necessity for the assistance of dermatologists, different systems have been proposed through the years with competing performances. Decision making in this case is of major importance because it would affect human life. In particular, content-based skin lesion retrieval systems have been recently investigated in order to explain the classification decisions and to increase trust by producing similar skin lesion images to a query ones. This property helps dermatologists understand the system’s decisions, and provides support to less experienced practitioners. In this study, we proposed an exploration of different similarity measures in function of different lesion-shaped expansion sizes. The main goal is to improve the performance of the deep learning-based skin lesion retrieval system by extracting the most relevant features from the skin lesion segment and its surrounding tissues, while selecting the appropriate similarity measure in this case. Moreover, within the framework of eXplainable Artificial Intelligence (XAI) and in order to further analyse and validate the proposed method, we employ attention maps. These maps highlight the features deemed relevant by the Inception-ResNet-v2 classifier for each test case, and this would make it easier to compare visually the different results. The obtained results on the challenging HAM dataset have indicated that the proposed method produces meaningful interpretations, which can effectively increase the accuracy of the content-based skin lesion retrieval.