Utility of Same-Modality, Cross-Domain Transfer Learning for Malignant Bone Tumor Detection on Radiographs: A Multi-Faceted Performance Comparison with a Scratch-Trained Model

Background/Objectives: Developing high-performance artificial intelligence (AI) models for rare diseases like malignant bone tumors is limited by scarce annotated data. This study evaluates same-modality cross-domain transfer learning by comparing an AI model pretrained on chest ra-diographs with a model trained from scratch for detecting malignant bone tumors on knee radio-graphs. Methods: Two YOLO-based models were developed: YOLO-TL, fine-tuned from a publicly available chest X-ray model, and YOLO-SC, trained from scratch using only our bone tumor da-taset. Both were trained and validated on the same institutional data and evaluated on an inde-pendent external test set of 743 knee radiographs (268 malignant, 475 normal). Performance was measured via area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and positive predictive value (PPV), focusing on a clinically critical high-sensitivity threshold (sensitivity ≥ 0.90). Results: Overall diagnostic performance was high and comparable, with no significant difference in AUC (YOLO-TL, 0.954 vs. YOLO-SC, 0.961; p=0.53). At the high-sensitivity threshold, both models achieved 0.903 sensitivity, but YOLO-TL showed higher specificity (0.903 vs. 0.867; p=0.037) and PPV (0.840 vs. 0.793; p=0.030) than YOLO-SC. Conclusions: Transfer learning may not improve overall AUC but can enhance practical performance at clinically crucial thresholds. By maintaining high detection rates while reducing false positives, the transfer-learning model offers superior clinical utility. Same-modality cross-domain transfer learning is an efficient strategy for developing robust AI systems for rare diseases, supporting tools more readily acceptable in real-world screening workflows.