Combining global gate axial-attention with U-Net for skin lesion segmentation

In dermatology, the mortality rate for melanoma is as high as 1.62%, imposing a significant burden on patients. Diagnosis primarily relies on the experience of physicians, who examine the skin’s color and texture. Traditional computer-assisted lesion segmentation has a low accuracy. However, with advancements in deep learning and the integration of various factors from medical imaging, the accuracy of lesion segmentation can be enhanced. Our skin disease detection method, based on deep learning, uses the GGAD-Net model and the SeDice loss function, with a weighting parameter of 0.5 for SeDice. On the ISIC 2018 dataset, after 50 epochs of training, the GGAD-Net model achieved an accuracy of 92.2%, a recall rate of 96%, a Dice score of 92.3%, and an IoU of 88.4%. The results prove that this model can effectively improve the performance of lesion segmentation, and in the future, we aim to further enhance accuracy and expand to other medical imaging segmentation applications.