Abstract
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 radiographs with a model trained from scratch for detecting malignant bone tumors on knee radiographs. METHODS: Two YOLOv5-based detectors differed only in initialization (transfer vs. scratch). Both were trained/validated on institutional data and tested on an independent external set of 743 radiographs (268 malignant, 475 normal). The primary outcome was AUC; prespecified operating points were high-sensitivity (≥0.90), high-specificity (≥0.90), and Youden-optimal. Secondary analyses included PR/F1, calibration (Brier, slope), and decision curve analysis (DCA). RESULTS: AUC was similar (YOLO-TL 0.954 [95% CI 0.937-0.970] vs. YOLO-SC 0.961 [0.948-0.973]; DeLong p = 0.53). At the high-sensitivity point (both sensitivity = 0.903), YOLO-TL achieved higher specificity (0.903 vs. 0.867; McNemar p = 0.037) and PPV (0.840 vs. 0.793; bootstrap p = 0.030), reducing ~17 false positives among 475 negatives. At the high-specificity point (~0.902-0.903 for both), YOLO-TL showed higher sensitivity (0.798 vs. 0.764; p = 0.0077). At the Youden-optimal point, sensitivity favored YOLO-TL (0.914 vs. 0.892; p = 0.041) with a non-significant specificity difference. 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.