Edge-Aware and Attention-Aided U-Net Model for Skin Lesion Segmentation

Segmenting skin lesions is essential for the early diagnosis of skin cancer, which significantly impacts therapy approaches and patient outcomes. Since lesion intensities and resolution dependencies vary, appropriate segmentation continues to be difficult to attain despite technological advancements. The primary objective of this research is to create a compact U-Net model that improves the segmentation performance by utilizing the Bit-Plane Attention technique and Effective Edge Detection technique in conjunction with the MobileNetV2 encoder. The suggested methodology outperforms the state-of-the-art models by managing non-linearity well and enhancing segmentation across various skin lesion types. The model achieves a 0.9296 Dice score, 0.9023 IoU score for the ISIC 2016 Task 1 dataset, and 0.9590 Dice score, 0.9230 IoU score for the PH2 dataset. The model is also tested on a histopathology dataset, called TNBC, to ensure its robustness, and it yields 0.8596 Dice score and 0.765 IoU score, which are better than existing approaches. The code of the proposed model is available at: https://github.com/Cmatermedicalimageanalysis/EAUNet.