A deep learning–based segmentation and CT–MRI fusion framework for preoperative spatial assessment of cervical ossification of the posterior longitudinal ligament

Technical Summary: Deep Learning–Based Segmentation and CT–MRI Fusion for Cervical OPLL

Problem Statement
Cervical ossification of the posterior longitudinal ligament (OPLL) is a degenerative condition causing spinal cord compression and cervical myelopathy, particularly prevalent in Asian populations. Accurate preoperative assessment requires evaluating the spatial relationship between the ossified lesion and the spinal cord. Currently, this relies on Computed Tomography (CT) for bony morphology and Magnetic Resonance Imaging (MRI) for soft tissue/spinal cord visualization. However, clinical practice faces two primary limitations:

1. Manual Delineation: Identifying and outlining OPLL lesions on CT is time-consuming and suffers from inter-observer variability.
2. Lack of Spatial Integration: CT and MRI are typically interpreted separately in 2D slices. The absence of a unified spatial framework hinders the accurate assessment of the three-dimensional (3D) relationship between the ossified mass and the spinal cord, complicating surgical planning.

Methodology
The study proposes a framework combining deep learning-based automatic segmentation with CT–MRI multimodal fusion.

• Data Acquisition: A retrospective cohort of 193 patients with cervical OPLL was analyzed. 173 cases were used for model training, and 20 served as an internal test set. Additionally, 10 matched patients with both CT and MRI scans (within one week) were selected for multimodal fusion analysis.
• Automatic Segmentation (CT):
  • Architecture: The backbone network is nnUNetv2, chosen for its automated configuration and adaptability.
  • Optimization: A Channel- and Spatial-Attention Module (CABM) was integrated into the encoder–decoder architecture. This module applies adaptive weighting to emphasize features relevant to target regions, specifically improving sensitivity to OPLL boundaries and irregular morphologies.
  • Task: The model performs multi-class segmentation of vertebrae, intervertebral discs, and OPLL.
  • Ground Truth: Annotations were manually delineated by two spine surgeons on axial CT slices, excluding anterior longitudinal ligament calcification and osteophytes, and reviewed by an experienced radiologist.
• Spinal Cord Extraction (MRI):
  • The open-source model TotalSpineSeg was utilized to extract spinal cord structures from MRI data without additional training.
• 3D Registration and Fusion:
  • Pipeline: A Python-based pipeline using SimpleITK and Open3D.
  • Process: Segmented vertebrae and discs from both CT and MRI were converted into 3D point clouds. A rigid registration was performed using the Iterative Closest Point (ICP) algorithm to align the MRI space to the CT space.
  • Fusion: The spinal cord mask from MRI was resampled into the CT coordinate system. Isosurfaces were generated using the Marching Cubes algorithm to create a unified 3D STL model containing ossified lesions, vertebrae, discs, and the spinal cord.
• Evaluation Metrics: Performance was assessed using the Dice Similarity Coefficient (DSC), Intersection over Union (IoU), 95th percentile Hausdorff Distance (HD95), and Average Symmetric Surface Distance (ASSD).

Key Results

• Segmentation Performance:
  • In five-fold cross-validation, the model achieved mean Dice scores of 0.98 for vertebrae, 0.91 for intervertebral discs, and 0.81 for OPLL (mean Dice: 0.90).
  • On the internal test set, the model achieved a mean Dice of 0.91, IoU of 0.85, and HD95 of 0.89 mm, indicating sub-millimeter boundary accuracy.
  • Manual annotation consistency (Kappa coefficient) between the annotator and reviewer was 0.89.
• Multimodal Fusion Performance:
  • After registration, the Dice and IoU for reference anatomical structures (vertebrae/discs) between CT and MRI were 0.9063 and 0.8286, respectively.
  • The ASSD was 0.8354 mm, and the HD95 was 4.1231 mm.
• Visualization: The fused model successfully visualized the spatial relationship between OPLL and the spinal cord in a unified coordinate space, allowing for multi-angle 3D inspection of contact and compression.

Significance and Claims
The authors claim that this study presents a technical framework that addresses the limitations of single-modality imaging analysis.

• Efficiency and Consistency: The deep learning approach improves the efficiency and consistency of image analysis compared to manual delineation, reducing subjectivity.
• Spatial Assessment: By integrating CT-derived osseous structures with MRI-based soft tissue information, the method enables an intuitive 3D visualization of the complex spatial relationship between ossified lesions and the spinal cord.
• Clinical Utility: The authors suggest the method may assist in preoperative spatial assessment and surgical planning, specifically for procedures like Anterior Controllable Antedisplacement and Fusion (ACAF), by helping surgeons visualize contact relationships and estimate decompression coverage.
• Decision Support: The study positions the tool as a means to facilitate understanding of 3D anatomical relationships to support clinical decision-making, rather than to directly replace it.

Limitations Acknowledged by the Authors

• Segmentation Accuracy: OPLL segmentation (Dice 0.81) was lower than that of vertebrae and discs, attributed to irregular morphology and indistinct boundaries.
• Registration Discrepancies: While overall alignment was good, localized deviations (reflected in HD95) persist due to inherent differences in imaging mechanisms and resolution between CT and MRI, particularly at soft tissue–bone interfaces.
• Scope: The current model focuses on spatial visualization and does not yet incorporate quantitative metrics (e.g., compression ratios, Cobb angle) for objective assessment.
• Generalizability: As a single-center retrospective study with a limited sample size, the generalizability of the model may be affected.