EvidSpine-Fusion: Registration-Free Cross-Modal State Space Fusion with Evidential Uncertainty for Trustworthy Early Detection of Spinal Cord Pathologies on Multi-Modal MRI and CT
DOI:
https://doi.org/10.51483/IJAIML.6.10s.2026.917-930Keywords:
AIOps; root cause analysis; microservices; cloud computing; observability; causal inference; multimodal learning; incident response; explainable AI; human-in-the-loopAbstract
Background: Early detection of spinal cord pathology is limited because no single modality is sufficient: magnetic resonance imaging (MRI) resolves the cord and intramedullary signal change, whereas computed tomography (CT) characterizes the osseous compression and ossification that mechanically drive it.
Gap: Existing fusion pipelines depend on accurate MRI–CT registration and emit over-confident, uncalibrated predictions, which restrict safe clinical use.
Objective: We aimed to fuse complementary neural and osseous evidence without explicit registration, at tractable cost, and with calibrated uncertainty suitable for triage.
Method: EvidSpine-Fusion couples dual cross-modal selective state-space (CM-S4) encoders, a deformable cross-modal fusion (DCMF) module that learns voxel-wise sampling offsets to align soft-tissue and osseous features without registration, and an evidential Dirichlet head trained with a cross-modal consistency objective. We developed the model on 1,824 paired cervico thoracic MRI–CT studies from four centre’s using patient-level five-fold cross-validation, with an independent external test set (n = 182).
Results: On external testing, EvidSpine-Fusion reached an AUC of 0.961 (95% CI 0.948–0.973), sensitivity 0.931, specificity 0.908, and lesion Dice 0.842, exceeding the strongest single-modality baseline by 6.8 AUC points (DeLong p < 0.001). Expected calibration error fell from 0.058 to 0.021, and deferring the most uncertain 12% of cases raised retained accuracy to 0.961. Under ±5 mm/±5° misalignment, accuracy fell 1.3 points versus 6.3 for concatenation fusion.
Significance: Registration-free cross-modal state-space fusion with evidential uncertainty improves early detection while remaining calibrated and robust to misregistration, offering a practical basis for trustworthy MRI–CT triage where the modalities are rarely co- registered.
