Deep Learning for Proximal Femoral Fracture Radiographs Under Archive-Label Uncertainty: A Duplicate-Controlled Single-Center Study

Background: Performance estimates in medical-imaging artificial intelligence may depend not only on model architecture but also on dataset integrity, label reproducibility, and the complexity of the prediction target. This study evaluated deep learning for extracapsular proximal femoral fracture radiographs within a duplicate-controlled, patient-level framework, comparing performance for fracture detection with prediction of fine-grained archive-derived AO/OTA group labels and quantifying the reproducibility of the historical labels through blinded reader reassessment. Methods: A retrospective single-center image archive was audited for exact and near-duplicate images, postoperative hardware, technical limitations, patient-level overlap, and label inconsistency. The final analytical cohort comprised 516 radiographs from 515 patient groups, including 75 no-fracture radiographs and 441 radiographs carrying locked archive-derived AO/OTA 31-A1, 31-A2, or 31-A3 analytical labels. A compact convolutional neural network, DenseNet121, ResNet50V2, and EfficientNetB0 were evaluated using three repetitions of five-fold stratified group cross-validation. Performance was estimated from averaged out-of-fold probabilities with 2000 patient-group bootstrap resamples. A second orthopedic trauma specialist subsequently performed a blinded reassessment to quantify label reproducibility. Grad-CAM was used for qualitative error analysis. Results: For the prespecified three-class archive-label prediction task, ResNet50V2 achieved the highest balanced accuracy of 0.395 (95% CI, 0.349–0.444), without a statistically significant advantage over DenseNet121. By contrast, for fracture-versus-no-fracture discrimination, EfficientNetB0 achieved a balanced accuracy of 0.890 (95% CI, 0.845–0.931) and a ROC-AUC of 0.964 (95% CI, 0.947–0.978). Blinded reassessment showed 99.8% agreement for fracture presence (Cohen’s κ = 0.992; 95% CI, 0.973–1.000) and 82.2% agreement for AO/OTA 31-A1/A2/A3 assignment (κ = 0.724; 95% CI, 0.666–0.782), with most subtype disagreements occurring between 31-A1 and 31-A2. Grad-CAM maps showed not only activation overlapping the fracture-side proximal femoral region in some cases but also broad or off-target activation. These qualitative findings do not establish that model predictions relied on clinically relevant fracture features. Conclusions: Deep learning showed strong internal discrimination between fracture and no-fracture radiographs within the reconstructed dataset, although systematic differences in archived image dimensions between these categories may have provided non-anatomical discriminative cues; performance in reproducing fine-grained historical AO/OTA group labels remained limited. Although subtype-label reproducibility was imperfect, the exploratory agreement-stratified analysis did not demonstrate systematically poorer classification performance in cases with discordant annotations. Therefore, the observed performance contrast cannot be attributed causally to annotation disagreement. The subtype results quantify performance against historical analytical labels rather than diagnostic accuracy against an adjudicated clinical consensus standard.