Mixed-source-to-institutional hospital-holdout evaluation of DICOM-aware adaptive harmonization for chest radiograph lesion detection

Robust cross-hospital generalization remains a major challenge in chest radiograph lesion detection, especially when detectors are trained on mixed public and institutional development sources and then transferred to a previously unseen hospital. This study evaluated whether a Digital Imaging and Communications in Medicine (DICOM)-aware adaptive harmonization branch, Adaptive HyFusion, changes held-out institutional lesion-detection performance under a mixed-source-to-institutional hospital-holdout design compared with standard preprocessing. A fixed Faster R-CNN detector with a ResNet-50-FPN backbone was evaluated under a harmonized five-class lesion ontology using VinDr-CXR, MS-CXR, and two non-public Indonesian hospital cohorts. Scenario 2A used RSUD Dr. Adhyatma, MPH Jawa Tengah for institutional development and RSUD Sunan Kalijaga Kabupaten Demak as the locked held-out test source; Scenario 2B reversed the hospital roles as a sensitivity analysis. The experimental protocol used patient/study-clean institutional development splits while preserving the locked held-out test sets. Performance was assessed using mAP 50 , mAP 50 : 95 , class-wise AP, recall, precision, false positives per image, threshold sensitivity, FROC-oriented summaries, and paired bootstrap uncertainty estimates. In Scenario 2A, Adaptive HyFusion showed numerically higher average held-out mAP 50 (0.127 ± 0.026 vs. 0.110 ± 0.073) and mAP 50 : 95 (0.042 ± 0.014 vs. 0.038 ± 0.023), but the paired bootstrap mAP 50 difference was small and uncertain (＋0.017; 95% CI: −0.052 to ＋0.099). In Scenario 2B, Adaptive HyFusion showed comparable mAP 50 to baseline (0.114 ± 0.011 vs. 0.117 ± 0.015) and lower false positives per image (0.985 ± 0.080 vs. 1.137 ± 0.114), but lower recall. Component ablations showed that isolated spatial, frequency, fixed-fusion, and adaptive-without-QC variants did not consistently improve held-out institutional performance. These findings support Adaptive HyFusion as an auditable DICOM-aware harmonization strategy that can alter held-out operating behavior under source shift, but they do not establish a statistically or operationally convincing generalization benefit on the primary Scenario 2A benchmark.