Deep learning-based segmentation and quantitative analysis of lumbar paraspinal muscles and fat infiltration in multicenter T2-weighted MRI data

This is the authors' abstract. We don't add key points for this paper.

Objectives The aim of this study was to develop deep learning algorithms based on T2WI to accurately segment the multifidus (MF), erector spinae (ES), and psoas major (PS) muscles, and to automatically calculate the degree of fat infiltration. Methods A total of 1457 participants with T2WI were enrolled from three hospitals. Models were trained and validated using data from the first hospital to segment the MF, ES, and PS. The data from other two hospitals were used for independently testing. We compared 2D and 3D segmentation models, using Dice and IoU as the main evaluation indicators. Intramuscular fat was segmented using the maximum inter-class variance method (OTSU). The consistency between the OTSU-based fat infiltration results and radiologist’s was evaluated using the weighted Kappa test. Results All trained models demonstrated high accuracy for segmenting MF, ES and PS (Dice ≥0.9) in the validation and testing cohorts. Overall, 2D models outperformed 3D models (p < 0.001), and nnUnet models outperformed Vnet and SEVB-Net (p < 0.001) across all muscles. The fat infiltration, assessed by OTSU, showed a high level of consistency with the radiologist. The consistency of MF, ES and PS at the patient level was 0.703, 0.751 and 0.838. At the slice level, the corresponding values were 0.619, 0.762, and 0.663. Conclusions The nnUNet segmentation network, which is based on 2D images, achieved the highest level of segmentation performance in MF, ES, and PS. The degree of fat infiltration, determined using OTSU, was highly consistent with that of a senior radiologist.