Skin lesion image segmentation based on improved TransUNet

For dermatologists to diagnose skin diseases, accurate skin lesion image segmentation is of great significance. Researchers are currently trying to use various deep learning models to complete this segmentation task. In skin lesion images, lesions differ widely in size and shape , which significantly affects how well these models segment the lesions. To better cope with these differences, we propose an improved TransUNet model based on DCNv3 and Triplet Attention (DTA-TransUNet) for skin lesion image segmentation in our paper. This model introduces deformable convolution DCNv3 into the encoder and decoder of TransUNet, and adds Triplet Attention module at the skip connection. We conduct experiments on ISIC2017, a dataset suitable for segmenting skin lesions in images, to compare our model with U-Net and TransUNet. By analyzing the experimental results, we see that DTA-TransUNet can better adapt to lesions of different sizes and shapes and achieve better segmentation results.