Attention-based deep learning in forensic anthropology: Validation of YOLOv12 architecture for automated skeletal fracture detection

To develop and validate a YOLOv12-based automated fracture detection system for forensic radiographic analysis, with particular focus on ante-mortem/post-mortem (AM/PM) comparison workflows where high specificity is critical.

In a retrospective validation study, 11,517 radiographic images depicting skeletal regions with and without fractures were obtained from Roboflow Universe. The dataset was split into training (n = 6908; 60%), validation (n = 2303; 20%), and test (n = 2306; 20%) subsets. YOLOv12-S (9.1 M parameters, 21.4 GFLOPs) was trained using five-fold stratified cross-validation (100 epochs per fold) followed by fine-tuning (50 epochs). Performance was evaluated with 1,000-iteration bootstrap (95% Confidence Interval) for mAP, sensitivity, specificity, PPV, and ROC-AUC.

YOLOv12-S demonstrated high stability in detection (mAP@0.5 =0.895, CV=4.80%). Image-level classification showed sensitivity of 86.50% (95% CI: 84.27%-88.74%) and specificity of 98.61% (95% CI: 97.80%-99.40%). ROC-AUC was 0.929 (95% CI: 0.918-0.940) with PPV of 98.33% (95% CI: 97.37%-99.27%). The Shapiro-Wilk test confirmed normal distribution of cross-validation metrics (W=0.873, p = 0.279), and Cohen's d of 1.38 indicated large effect size relative to baseline. Single-image inference time was 20.6 ms, enabling real-time throughput at 48.5 FPS.

The YOLOv12-S model shows promising potential for forensic radiographic screening due to its high throughput and stability (CV = 0.77%). Its particularly high specificity (98.61%) minimizes false positive identifications, which is critical in AM/PM comparison where erroneous fracture detection could lead to incorrect identity conclusions. While the model demonstrates strong technical performance, further validation in real forensic environments and comparison with expert examiners are required before it can be considered for operational use.