Vision Transformer-Based Automated Ordinal Grading of Supraspinatus Tendon Pathology and a Quantitative Faithfulness Audit of Attribution Maps in Shoulder MRI.

Authors

Abstract

Rotator cuff disorders, most often of the supraspinatus tendon, are a leading cause of shoulder pain, and magnetic resonance imaging is the reference modality. Grading their progression is subject to inter-reader variability, and automated systems for the task are expected to justify themselves through attribution maps. This study asks whether they do. Four Vision Transformer tiers spanning a 54-fold parameter range were trained under one searched configuration and averaged into an unweighted ensemble. On 694 regionally annotated sections from 276 shoulders in 264 patients, the ensemble reached 0.9130 accuracy (95% CI, 0.8406-0.9710) and a 0.9431 quadratic weighted kappa on a 69-section, patient-level held-out test subset. The four tiers differed by at most 0.0580, inside the 0.07 resolution limit, so any capacity effect is bounded below it rather than shown absent. Seven attribution methods from four mechanism families were scored against an occlusion reference. Occlusion placed 16.82 times chance mass on the annotated region, the attribution methods 1.15 to 13.90. Against a random ordering of the same cells, occlusion separated the deletion and insertion curves by 0.4773, a margin no method reached, and two methods beat this baseline on deletion but lost on insertion. The method closest to the reference kept a correlation of 0.977 with its own map after half the network was randomized. Repeating the audit under a Gaussian blur fill left the ranking of all eight arms unchanged. An attribution map offered as clinical evidence therefore requires interventional validation, with its perturbation and grid sensitivity reported.

Tags

Topics