Dcctnet: Kidney Tumors Segmentation Based On Dual-Level Combination Of Cnn And Transformer

Abstract

Translate Article

The hybrid model of CNN(Convolution Neural Networks) and Transformer is a popular method in segmenting kidney images, but most existing hybrid models directly fused local features from CNN with global features from Transformer, ignoring the issue of semantic gaps between distinct features. Furthermore, feature fusion is typically performed solely at the feature level, without considering alignment at the mask (prediction map) level. To address these limitations, we propose a novel segmentation method called Dual-level Combination of CNN and Transformers Network (DCCTNet). Specifically, we select similar features from both CNN and Transformer to reduce semantic gaps at the feature level. Additionally, we further utilize the global information of the Transformers by reducing the difference between the prediction maps in the coding stage at the mask level. We evaluate DCCTNet on the KiTS19 dataset, achieving $97.3 \%$ dice score for kidneys segmentation and $81.2 \%$ dice score for kidney tumors segmentation, respectively. https://github.com/hou-bz/DCCTNet.