DP-U-Net plus plus : inter-layer feature fusion for colorectal gland image segmentation

Colorectal adenocarcinoma is a common malignant digestive tract tumor from the colonic gland epithelium. The sliced microscopic images of the colorectal gland have great diversity in shapes, large differences in scale, and it is difficult for traditional networks to segment the glandular regions efficiently. In this paper, we propose a novel inter-layer feature fusion network named DP-U-Net++, which introduces improved deep feature fusion and enhanced dense skip-connection, and using pretrained ResNet-34 to extracts features. To combine with the advantages of Self Attention, position embedding is incorporated for correlation computation, and channel relevance in each layer is enhanced by an extra ECA block. Inspired by Attention U-Net, we propose a new structure called Dilated Spatial Attention Gate which expands the receptive field in feature fusion. Finally, the output features of the decoder derived from different branches are fused with low-level spatial features to achieve better performance. we conducted experiments on the GlaS and CRAG datasets and a verification experiment on the PH2 dataset. Compared with the baseline, our method improves metrics by 1.61% and 1.27% in dice coefficient and by 2.37% and 3.22% in mIOU on GlaS and CRAG, respectively. Experiments on gland segmentation datasets demonstrate that the proposed method in this paper outperforms baselines. Our source code can be found at https://github.com/icm162/ModelFrame.