Improved Skin Lesion Segmentation and Classification using Hierarchical Deep Attention Architectures

The challenges posed by skin diseases, including cancers, have led to growing interest in automating dermatological diagnostics using machine vision and deep learning. Models such as UNet and its derivatives have shown their effectiveness in the segmentation of skin lesions. In this article, we introduce an automated deep learning system designed to hierarchically segment and classify skin lesions, while providing the user with information useful for interpreting the classification results. Our study focuses first on evaluating the performance of a combined Attention module with UNet model «Att-UNet» to detect pertinent regions within skin lesions against baseline models, and makes classification decisions using different networks such as CNN, VGG16 and Resnet50. Evaluated on the ISIC17 and ISIC16 databases, the results highlight the specific advantages of each model, in particular the positive impact of the integration of attention mechanisms. All models exceeded a threshold of 90% precision, with a Dice coefficient of 0.953 on the validation set and 0.966 for the test set for Attention-UNet.