MpVit-Unet: Multi-path Vision Transformer Unet for Sellar Region Lesions Segmentation

The sellar area is located in the central part of the middle cranial fossa and the surrounding area with a complex structure, which is a common site of intracranial diseases. At present, the differential diagnosis of parasellar lesions mainly relies on magnetic resonance imaging, but the imaging features are very similar, and the huge differences in tumor size, shape, and appearance make the task of segmentation of sellar lesions still challenging. Due to the inherent limitations of convolution operations, Methods based on convolutional neural networks cannot meet the requirements of segmentation accuracy in the saddle region lesion segmentation task. The proposal of Vision Transformer (Vit) effectively solves these limitations and can efficiently learn global and long-range information interaction. Inspired by the Multi-Path Vision Transformer (MpVit), We propose a U-shaped encoder-decoder structure using MpVit as the backbone network and propose a skip connection with ResBlock, which we call MpVit-Unet. Experiments show that our proposed model achieves 1.45% and 0.8 % higher than the state-of-the-art methods on the T2W1 dataset mIou score and Dice score. On the CE-T1WI dataset, the mIou score and Dice score achieve 1.64% and 0.54% higher than the state-of-the-art methods.