Enhanced Skin Lesion Segmentation: DeepLabV3and U-Net with Spatial Attention Mechanisms

Accurate segmentation of skin lesions is crucial for early skin cancer detection and treatment. This study introduces a novel approach that enhances DeepLabv3 and U-N et architectures with spatial attention mechanisms to improve skin lesion segmentation. Using the ISIC dataset, the study shows that integrating attention mechanisms into these models significantly enhances their performance. DeepLabV3captures multi-scale contextual information with its atrous spatial pyramid pooling (ASPP), while U-Net excels at capturing fine-grained details through its encoder-decoder structure. The spatial attention mechanisms enable the models to focus dynamically on the most relevant image regions, improving accuracy and robustness. The models are trained on the diverse ISIC dataset, using data augmentation techniques like rotation, scaling, and color jittering to handle variability in lesion appearance and limited data size. A composite loss function balancing pixel-wise accuracy and boundary precision guides the training process. Experimental results demonstrate that the attention-enhanced DeepLabV3and U-Net models outperform their baseline versions achieving higher accuracy and Intersection over Union (IoU) scores. This study highlights the potential of attention-enhanced DeepLabV3 and U-Net models for skin lesion segmentation, suggesting they could be valuable tools for dermatologists in early skin cancer detection and treatment. The proposed models not only improve segmentation accuracy but also offer scalable solutions for clinical applications.