A biomimetically inspired staged pipeline for primary bone tumor detection on radiographs: held-out FROC evaluation under explicit false-positive constraints

Primary bone tumors are rare and radiographically subtle, and summary metrics such as the area under the receiver operating characteristic curve (AUC) do not constrain how many false marks a detection system places per examination. We evaluated radiograph-based detection with free-response receiver operating characteristic (FROC) analysis, the established endpoint for computer-aided detection, at explicit false-mark budgets. The system comprises two independently trained components whose organization is loosely inspired by expert reading: an image-level classifier (EfficientNetV2-S pretrained on MURA and fine-tuned on BTXRD with a two-phase unfreezing schedule) and a YOLOv8 lesion detector initialized from COCO weights; the classifier can optionally gate the detector. The 3,746 BTXRD radiographs (1,867 lesion-positive, 1,879 normal; three sources) were partitioned into training, validation and held-out test sets (2,265/738/743 images), and every threshold was fixed on validation data before the test set was analysed. Matched ablations compared MURA with ImageNet initialization and two-phase with full fine-tuning, and leave-one-center-out (LOCO) experiments assessed transfer across sources. On the test set, malignant-versus-normal AUC was 0.933 (95% CI 0.900-0.966) with MURA initialization and 0.910 (0.872–0.947) with ImageNet initialization; the effect of musculoskeletal pretraining was consistent in direction across three seeds (+ 0.023 to + 0.073), suggestive of a modest benefit, with the interval including zero at the pre-specified seed. At an overlap criterion of IoU ≥ 0.3, lesion-level sensitivity was 0.646 (0.602–0.690) at an achieved 0.027 false marks per normal radiograph (validation budget 0.02) and 0.698 (0.655–0.741) at an achieved 0.055 (validation budget 0.05), with 0.40 false marks per lesion-positive radiograph at the primary operating point. Sensitivity was 0.736 for malignant and 0.631 for benign lesions. Under LOCO it ranged from 0.669, when a small source was held out, to 0.287, when the largest source was held out and only 808 training images remained. Sensitivity at fixed false-mark budgets conveyed information that curve-level summaries did not; cross-source transfer and clinical utility at realistic prevalence remain to be established.