Enhancing Skin Lesions Segmentation and Classification through Hierarchical Attention ResUnet Model

Significant advances in the field of digital dermatology are due to the challenges posed by skin diseases, particularly cancers. Automating dermatological diagnostics through machine vision and deep learning has proven to be a promising approach. In this study, we present an automatic deep learning system dedicated to hierarchical segmentation and classification of skin lesions. Our approach integrates an Attention module combined with the ResUNet model, named “Att-ResUNet”. Our main contribution lies in the integration of an attention module combined with the ResUNet model, making it possible to precisely identify relevant areas within skin lesions, thus surpassing the performance of basic models. Furthermore, our method guides classification decisions by leveraging various networks such as CNN, VGG16, and Resnet50. Our evaluation of the ISIC17 and ISIC16 databases highlights the specific advantages of each model, particularly the positive influence of integrating attention mechanisms and residual blocks. All models show good performance, exceeding a precision threshold of 90%, with a Dice coefficient reaching 0.971 for the validation set and 0.981 for the test set, for the Attention-ResUNet model.