From Adults to Children: Benchmarking Foundation Models for Chest X-Ray Classification under Domain Shift

Foundation models are increasingly used as image feature extractors for mammography, but their robustness under external domain shift remains unclear. We benchmark 15 foundation-model backbones across breast density, BI-RADS severity, and cancer status using a unified frozen-backbone linear-probe protocol, training on 3 source datasets and evaluating on 12 task-compatible out-of-distribution (OOD) datasets after label harmonization. Mammography-specific vision-language models (Mammo-FM and MaMA) provide the strongest mean OOD performance, but robustness is not explained by mammography exposure alone. DINOv3 remains a competitive vision-only baseline, and mammography-adapted pretraining does not consistently improve generalization. Dataset-level analysis further shows that even leading models show heterogeneous performance across datasets. Feature-space inspection reveals that useful representations can preserve clinical signal while retaining dataset and acquisition structure. These findings highlight dataset-level OOD evaluation as a central criterion for assessing mammography representations. Our code is publicly available: https://github.com/biomedia-mira/mammo-ood.

Background and Objective: External evaluation of medical-imaging AI is often collapsed into discrimination. We evaluated a computational protocol that separately tests discrimination, probability calibration, fixed operatingpoint transport, shortcut-associated signal, and limited-label recoverability for pediatric pneumonia classification across datasets from three countries. Methods: After exact-duplicate removal, 5,824 Guangzhou radiographs supported leakage-controlled source development and internal testing. A frozen three-seed DenseNet121 dual-view ensemble was evaluated zero-shot on BDCXR-3257 from Bangladesh (n = 3, 257) and an untouched harmonized VinDr-PCXR/PediCXR test cohort from Vietnam (n = 1, 077). Matched seed-42 variants tested architectural robustness. Secondary BDCXR analyses used a fixed 651-image adaptation pool and 2,606-image hold-out; 163, 326, and 651 labels represented 5%, 10%, and 20% of complete BDCXR. Results: Internal AUROC was 0.976 with 95.1% sensitivity. BDCXR and VinDr-PCXR AUROC were 0.798 and 0.742, while frozen-threshold sensitivity fell to 6.2% and 0%. Source-to-BDCXR AUROC degradation occurred for a full-image baseline (0.961 to 0.749), ungated dual-view model (0.977 to 0.766), and gated MixStyle model (0.966 to 0.789). With 163 BDCXR labels, Platt recalibration preserved AUROC while increasing held-out sensitivity to 88.3%, but specificity was 47.9% and the alert rate was 78.5%. Two hundred repeated 163-label fits confirmed sensitivity recovery but substantial specificity variability. Conclusions: Cross-dataset shifts across countries affected ranking, probability alignment, and source-defined decision behavior differently. Transport studies should evaluate these components separately and quantify the operational burden of apparent recovery.

Foundation models are increasingly adapted for downstream medical imaging tasks, yet the influence of the chosen adaptation strategy on subgroup fairness remains poorly understood. We investigate how three parameter-efficient adaptation techniques, including linear heads on the raw CLS token, an MLP, and an attention-pooling module over multi-layer patch features, affect both pathology classification performance and subgroup disparities when applied to the frozen Rad-DINO chest X-ray encoder. Using MIMIC-CXR, we evaluate eight pathologies across race, sex, and imaging-view subgroups on a prevalence-preserving, demographically balanced test set, and additionally probe how strongly each adapter encodes protected attributes. We find that attention pooling achieves the strongest overall discriminative performance and encodes attributes, particularly race, most strongly, but that improved overall performance does not consistently reduce subgroup disparities. Notably, stronger attribute encoding did not correspond to larger disparities: early network layers encoded race most weakly yet produced the largest subgroup performance gaps. Exploring different attention-pooling layer combinations further revealed no consistent relationship between the layers pooled, attribute encoding strength, and subgroup fairness. Our results indicate that richer, more expressive representations can improve accuracy while leaving fairness implications task-dependent and unpredictable, which must be assessed directly and per-task rather than inferred from encoding strength or overall performance alone.