MCUNeXt: An Efficient U-Shaped Network for Pathology Image Segmentation

Abstract

Accurate pathology image segmentation is a process that assists physicians in developing medical plans and evaluating the effectiveness of treatment. Due to the complex background of pathology images and many cell nuclei, manual segmentation is time-consuming and laborious, so it is important to design a model for automatic segmentation of pathology images. In this paper, we proposed a U-shaped network that can effectively improve the accuracy and reduce the model complexity. We use depth-separable convolution and convolution kernels of different sizes to admit the convolutional blocks of UNet with the aim of reducing the number of parameters, obtaining multi-scale capabilities, and being able to effectively combine global and local information. Also, the inverted bottleneck structure is proposed to be able to increase the accuracy while reducing the number of parameters. We replace pooling with convolution for downsampling and are able to retain more information. Meanwhile, our proposed method has been extensively experimented on CRAG dataset, and compared with the standard UNet, our parameter quantity is reduced by 36%, Jaccard is 3.61% higher, and our method has less hole phenomenon and sticking phenomenon, better boundary accuracy compared with other excellent methods, and the results show that our method is competitive.