Accurate Segmentation and Three-dimensional Reconstruction Algorithm of Spinal Cord Injury Lesions Based on Multimodal Magnetic Resonance Imaging.

INTRODUCTION

Accurate segmentation and Three-Dimensional (3D) reconstruction of Spinal Cord Injury (SCI) lesions are essential for clinical diagnosis, surgical planning, and prognostic assessment. However, multimodal MRI-based SCI lesion analysis faces critical challenges, including blurred lesion boundaries, non-rigid registration errors between modalities, and frequent modality loss in real clinical scenarios. This study aims to develop a robust network to address these issues and improve the clinical applicability of automated SCI lesion analysis.

METHODS

A Cross-Modal Alignment and Uncertainty-Aware Network (CMUA-Net) was proposed. It integrated three core components: 1) a learnable Cross- Modal Attention Alignment Module (CMAM) to dynamically correct spatial misalignments of T1w, DTI-FA, and T2-star-weighted (T2*w) at the feature level using T2w as the semantic reference, avoiding traditional image-level registration interference; 2) a Multi-Scale Residual Dense Block (MS-RDB) to fuse convolutional features with multiple receptive fields, enhancing contextual modeling of small and diffuse lesions; 3) a Modality-Robust Training (MRT) strategy combining random modality masking and Monte Carlo dropout-based uncertainty-guided loss to improve adaptability to any modality combination. The network was validated on multimodal MRI data from 127 multicenter SCI patients.

RESULTS

With quad-modal input, CMUA-Net achieved a Dice Similarity Coefficient (DSC) of 0.821 and a 95% Hausdorff Distance (HD) of 3.21 mm; even with only T2w input, the DSC remained 0.713. Its 3D reconstruction yielded a Fréchet Video Distance (FVD) of 18.7. The network outperformed baseline methods (Spinal Cord Toolbox, U-Net, nnU-Net, TransBTS) in all metrics (p < 0.01), and the R2 between its automated lesion volume measurement and manual physician measurements reached 0.998.

DISCUSSION: CMUA-Net effectively resolves clinical challenges such as strong subjectivity in quantification, incomplete modalities, and manual dependence in modeling. Its high multi-center adaptability and stability in small-lesion segmentation make it suitable for cross-institutional promotion, providing reliable anatomical references for prognosis assessment, doctor-patient communication, and rehabilitation monitoring.

CONCLUSION

CMUA-Net enables high-precision, robust SCI lesion segmentation and 3D reconstruction, and can automatically output lesion volume, 3D models, and key surgical parameters. It provides technical support for personalized SCI diagnosis and treatment, with significant clinical applicability and promotion value.