CPA-UNet: nnUNet-Based Local Pyramid Aggregation Network

In this paper, we propose CPA-UNet, a segmentation model for diabetic foot ulcers based on the nnUNet architecture. Initially, the model employs a ResNet-D module for diabetic foot ulcer images to establish residual connections, thereby reducing information loss in skip connections. Subsequently, a Local Pyramid Attention (LPA) module is utilized to aggregate both local and global attention, enhancing the focus on the lesion area, in addition, the accuracy is improved by incorporating the ConvNeXt convolution module. The results of our experiments using the DFUC2022 dataset indicate that, compared to the classic models Mask R-CNN, FCN, U-Net, DeeplabV3, and nnUNet, for diabetic foot ulcers, the proposed model achieved segmentation accuracy improvements of 6.5%, 5.7%, 7.3%, 3.7%, and 1.9%, respectively.