DCCUNet: A double cross-shaped network for pathology image segmentation

Cell nuclei offer valuable insights into the microenvironment, making automatic cell/nuclei segmentation crucial for quantitative pathological analysis. Despite the remarkable achievements of existing methods, accurate pathology image segmentation remains a challenge due to the presence of numerous cell clusters, high variability in appearances, tissue overlap, and complex backgrounds. In this work, we developed two cross-shaped modules and integrated them into the encoder and skip connections within the UNet architecture to achieve effective and robust segmentation of pathology images. Specifically, our approach incorporates a parallel asymmetric convolution module to extract hierarchical multi-scale features. This cross-shaped convolution module imposes a restriction on the convolution kernel, inducing the network to prioritize the image block center with a larger weight. Furthermore, we introduced a depthwise recurrent criss-cross attention mechanism within the skip connections to further emphasize the importance of the block center, resulting in more distinctive features. Extensive experiments demonstrate the strong generalization capabilities and competitive performance of our proposed model across various pathology image databases for cell segmentation. The ablation study validates the effectiveness and advantages of parallel asymmetric cross convolution module and depthwise recurrent criss-cross attention mechanism. The code is available at: https://github.com/zsj0577/DCCUNet.