A New Cine-MRI Segmentation Method of Tongue Dorsum for Postoperative Swallowing Function Analysis

Advantages of cine-MRI include high spatial-temporal resolution and free radiation, and the technique has become a new method for analyzing and assessing the swallowing function of patients with head and neck tumors. To reduce the labor work of physicians and improve the robustness of labeling the cine-MRI images, we propose a new swallowing analysis method based on a revised cine-MRI segmentation model. This method aims to automate the calculation of tongue dorsum motion parameters in the oral and pharyngeal phases of swallowing, followed by a quantitative analysis. Firstly, based on manually annotated swallowing structures, we propose a method for calculating tongue dorsum motion parameters, which enables the quantitative analysis of swallowing capability. Secondly, a spatial-temporal hybrid model composed of convolution and temporal transformer is proposed to extract the tongue dorsum mask sequence from a swallowing cycle MRI sequence. Finally, to fully exploit the advantages of cine-MRI, a Multi-head Temporal Self-Attention (MTSA) mechanism is introduced, which establishes connections among frames and enhances the segmentation results of individual frames. A Temporal Relative Positional Encoding (TRPE) is designed to incorporate the temporal information of different swallowing stages into the network, which enhances the network’s understanding of the swallowing process. Experimental results show that the proposed segmentation model achieves a 1.45% improvement in Dice Score compared to the state-of-the-art methods, and the interclass correlation coefficient (ICC) of the displacement data of swallowing feature points obtained respectively from the model mask and physician annotation exceeds 90%. Our code is available at: https://github.com/MinghaoSam/SwallowingFunctionAnalysis .