SCU-Net++: A Nested U-Net Based on Sharpening Filter and Channel Attention Mechanism

U-Net++ is one of the most prominent deep convolutional neural networks in field medical image segmentation after U-Net. However, semantic gaps between encoder and decoder subnets are still large, which will lead to fuzzy feature maps even target regions segmentation. To solve this problem, we propose an improved model utilizing channel attention mechanism Laplacian sharpening filter, SCU-Net++: dense skip connections redesigned with filters ease gaps, modules used make pay more on that useful for our pixel-level classification task. Compared U-Net++, proposed obtains a competitive performance Pancreas Segmentation dataset Liver Tumor dataset, while increases very small number learnable parameters thus almost does not additional training reasoning costs. The method carried out supervision mode, alleviates problem gradient disappearance, pruning can be activated accelerate speed.