AI Model Detects Skeletal Fractures With 98.61% Specificity in Forensic Study

Researchers at Ankara University in Türkiye have evaluated an artificial intelligence system designed to detect skeletal fractures in radiographic images, reporting 98.61% specificity under experimental conditions. The findings, published online on October 26, 2026, in Forensic Science International, suggest that automated image analysis could support forensic identification procedures.

Skeletal injuries can provide valuable information when forensic anthropologists compare medical records obtained before death with examinations conducted after death. These comparisons may help establish whether unidentified remains are consistent with a particular missing person.

However, identifying fractures in radiographic images requires careful interpretation. Anatomical variations, image quality, and overlapping structures can complicate examination.

Researchers Cumali Çatak and Mert Ocak investigated whether an attention-based deep learning architecture, YOLOv12, could assist with fracture detection.

The team used 11,517 radiographic images depicting skeletal regions with and without fractures, obtained from Roboflow Universe.

The dataset was divided into 6,908 training images, 2,303 validation images, and 2,306 test images.

Researchers trained the YOLOv12-S model using five-fold stratified cross-validation, followed by additional fine-tuning.

The system achieved a mean average precision of 0.895 at an intersection-over-union threshold of 0.5, indicating promising fracture localization performance.

For image-level classification, sensitivity reached 86.50%, while specificity was 98.61%.

Sensitivity measures the proportion of fracture-positive images correctly identified. Specificity measures the proportion of fracture-negative images correctly classified.

The high specificity is particularly relevant to forensic identification because incorrectly detecting a fracture could introduce misleading information into comparisons between ante-mortem and post-mortem records.

However, the reported sensitivity also indicates that the system missed some fracture-positive images during testing.

The researchers reported a receiver operating characteristic area under the curve of 0.929, reflecting the model's ability to distinguish images with and without fractures across classification thresholds.

The system processed individual images in approximately 20.6 milliseconds, equivalent to a reported throughput of 48.5 frames per second.

These results suggest that automated screening could help forensic specialists prioritize radiographic images for closer examination.

Nevertheless, detecting a fracture is not equivalent to identifying an individual. Establishing identity requires evaluation of the fracture's location, morphology, healing characteristics, and other supporting evidence.

The study also relied on an existing image dataset rather than demonstrating performance during routine forensic casework.

Independent testing on representative forensic radiographs and direct comparisons with qualified examiners remain necessary.

The research demonstrates the potential of AI-assisted skeletal examination while emphasizing that automated findings should support, rather than replace, expert forensic interpretation.