Opportunistic rotator cuff tear screening from routine chest CT: a deep learning-radiomics hybrid model with multi-center validation

Background Rotator cuff tears (RCT) constitute the predominant etiology of shoulder dysfunction among middle-aged and elderly populations yet remain substantially underdiagnosed in community settings and routine health screenings owing to the prohibitive cost of magnetic resonance imaging. Chest computed tomography (CT), as a ubiquitously accessible imaging modality in contemporary clinical practice, furnishes a natural data substrate for opportunistic RCT screening. Notwithstanding, real-world implementation confronts formidable challenges encompassing suboptimal soft tissue contrast, incomplete anatomical coverage, and pervasive equipment heterogeneity. Methods We developed a fully automated, opportunistic screening system validated across a diverse multi-center cohort (N = 1,442) from national, municipal, and county-level hospitals. To overcome imaging variations, we implemented a rigorous preprocessing pipeline incorporating ComBat harmonization to minimize cross-center batch effects. The core architecture features a gated attention-based multiple instance learning (Attn-MIL) network to extract deep representations from 3D CT patches. These were synergistically fused with interpretable radiomic features quantifying muscle compensation and osseous degeneration. Results In the primary national-center cohort, the hybrid model yielded excellent diagnostic discrimination (AUC, 0.956; 95% CI, 0.943–0.969). Across highly heterogeneous real-world external validation cohorts, the model exhibited robust generalizability: AUC attained 0.893 at the municipal center and 0.858 at the county center. Subgroup analyses confirmed model robustness across divergent body habitus, scanner manufacturers and acquisition protocols. Notably, the model maintained high precision in identifying early-stage pathology, validating its sensitivity for occult injury screening. Interpretability analyses further corroborated that the model correctly captured kinetic chain reorganization patterns secondary to RCT. Conclusions This study establishes the clinical feasibility of opportunistic RCT screening using routine chest CT. By effectively neutralizing batch effects and leveraging deep feature fusion, our system enables accurate, zero-cost risk stratification without additional radiation, facilitating proactive population health management.