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

Area of Science:

• Medical Imaging
• Neuroscience
• Artificial Intelligence

Background:

• Accurate segmentation and 3D reconstruction of spinal cord injury (SCI) lesions are crucial for clinical applications.
• Existing multimodal MRI-based SCI lesion analysis faces challenges like blurred boundaries, registration errors, and modality loss.

Purpose of the Study:

• To develop a robust network addressing the limitations of current SCI lesion analysis methods.
• To improve the clinical applicability of automated SCI lesion segmentation and 3D reconstruction.

Main Methods:

• Proposed a Cross-Modal Alignment and Uncertainty-Aware Network (CMUA-Net) integrating Cross-Modal Attention Alignment Module (CMAM), Multi-Scale Residual Dense Block (MS-RDB), and Modality-Robust Training (MRT).
• CMAM dynamically corrects spatial misalignments at the feature level using T2w as reference.
• MRT employs random modality masking and uncertainty-guided loss for adaptability to various input modalities.

Main Results:

• CMUA-Net achieved a Dice Similarity Coefficient (DSC) of 0.821 with quad-modal input and 0.713 with only T2w input.
• Demonstrated superior performance over baseline methods in all metrics (p < 0.01) and high correlation (R2=0.998) with manual lesion volume measurements.
• Achieved a Fréchet Video Distance (FVD) of 18.7 for 3D reconstruction.

Conclusions:

• CMUA-Net effectively overcomes subjectivity, modality incompleteness, and manual dependence in SCI lesion quantification.
• The network exhibits high multi-center adaptability and stability, particularly for small lesions, making it suitable for clinical promotion.
• Provides reliable anatomical references and technical support for personalized SCI diagnosis, treatment, and monitoring.