Deep Learning-Based Melanoma Detection UsingAttention Maps

Melanoma is a cancer type with notably high fatality rates whose prevalence has been increasing in the last two decades and is forecasted to be even more prevalent in the future. Consequently, the development of non-invasive and highly efficient diagnostic methods tailored to melanoma is urgently needed. Among the methods employed for diagnosis is the examination of dermoscopic images of skin lesions, which can also be analyzed via Deep Learning, as previous scholarly works have shown. We present an attention-based framework predicated upon the supposition that the lesion is the most informative part of dermoscopic images. As a result, the proposed method performs the task of lesion localization before classification. Subsequently, leveraging a Deep Learning-driven feature extractor, black-box features are obtained for the lesion segment, complemented by a feature extractor to extract features from the entire image. A composite feature vector is then obtained by concatenating the two feature vectors and, upon processing through a simple neural network, enables image classification. Empirical validation conducted on the ISIC 2019 dataset with the proposed methodology yielded an accuracy of 75.5%.