Towards Clinically Interpretable Skin Lesion Analysis: A Unified Deep Learning Framework for Classification and Localization
TL;DR
An explainable deep learning architecture for skin lesion classification and localization based on the HAM10000 dataset is proposed that integrates the concepts of classification representation learning and lesion localization in a single architecture by incorporating image-based lesion classification alongside lesion localization through segmentation masks.
Abstract
The analysis of skin lesions has been highlighted as an essential use of deep learning in the context of understanding medical images, with particular emphasis on their capacity for aiding early diagnosis and supporting dermatology screening. Nevertheless, there is a growing number of deep learning-based approaches in which there is little consideration for lesion localization and lack of interpretability in the prediction process. With this in mind, this paper proposes an explainable deep learning architecture for skin lesion classification and localization based on the HAM10000 dataset. Such framework integrates the concepts of classification representation learning and lesion localization in a single architecture by incorporating image-based lesion classification alongside lesion localization through segmentation masks. Moreover, in order to improve the inter¬pretability of predictions, the use of Grad-CAM visualization is incorporated. For the evaluation of lesion localization tasks, several segmentation architectures such as U-Net, FCN-ResNet50, DeepLabV3-ResNet50, and Attention U-Net are considered along with metrics of Intersection-over-Union and Dice coefficient. The results indicate that the proposed architecture is able to find a good trade-off between the three characteristics mentioned above. The paper addresses several challenges such as class imbalance and lack of external validation as well as dataset specific bias.... Impact Statement—The current research makes an important contribution towards clinical understanding through the integration of lesion recognition, localization of lesion regions, and interpretability within the framework. As compared to prediction-based approaches, the emphasis of lesion interpretation and practical limitations in the research will enhance the credibility of AI-driven dermatology applications. While the current study is based on a publicly available benchmark data set and experimentally controlled setting, it provides a basis for future developments of diagnosis systems