Ordinal Multi-Instance Learning for Automated Multi-Task Grading of Lumbar Degenerative Conditions on Sagittal MRI, with External Validation
Year: 2026
Type: preprint
Abstract: Purpose Lumbar spine MRI interpretation requires 25 structured severity judgments per examination — five degenerative conditions at five disc levels, each on a three-level ordinal scale — a time-consuming task with substantial inter-reader variability. We developed and externally validated a model that automates this complete scheme. Materials and Methods A full-sequence multi-instance learning (MIL) model — ResNet34 backbone, bidirectional GRU, attention pooling over 24 uniformly sampled sagittal T2-weighted slices — with 25 cumulative-logit ordinal heads was trained on 1,204 examinations from the multi-institutional RSNA 2024 LumbarDISC dataset and internally validated on 213 held-out examinations. External validation used 60 retrospectively collected patients from an independent Chinese university hospital (four scanner vendors, radiology reports as the reference standard, no fine-tuning). Agreement was assessed with exact accuracy, sensitivity, specificity, McNemar tests, and patient-clustered bootstrap confidence intervals. Results The ordinal model achieved 51.4% balanced accuracy internally (exact accuracy 65.1%), versus 16.3% for a single-slice baseline; severe-grade AUCs averaged 0.908 across conditions. Externally, it outperformed its softmax-weighted counterpart on exact accuracy (0.649 vs. 0.586; +6.3 percentage points, 95% CI +2.7 to +10.0) and sensitivity (0.868 vs. 0.833), with markedly fewer severe-grade false alarms (69 vs. 172 predicted severe cells; McNemar p<0.0001). Conclusion A full-sequence MIL model with ordinal outputs automates the complete 25-task grading scheme on sagittal MRI and, in the first hospital external validation of such a model, reduced severe-grade over-prediction under domain shift without loss of sensitivity — supporting its role as a screening and triage aid. Summary An ordinal full-sequence multi-instance learning model graded all 25 lumbar MRI degeneration tasks and markedly reduced severe-grade false alarms in 60-patient hospital validation without loss of sensitivity. Key Points Full-sequence attention was indispensable: balanced accuracy rose from 16.3% (single-slice baseline) to 51.4% over 24 sagittal slices. In external validation, ordinal outputs reduced severe-grade predictions from 172 to 69 of 900 cells (McNemar p<0.0001) with comparable ranking ability (AUC 0.835 vs. 0.829). The Moderate grade remained the hardest class across all conditions (AUC 0.70–0.78), tracking known inter-reader variability at the mild–moderate boundary.
Source: medRxiv
Institution: First Hospital of Lanzhou University
Language: English
Topic: Medical Imaging and Analysis
Subfield: Biomedical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7223246022