Automated Skull Fracture Detection in CT Imaging: A Systematic Review of Computational Methods, Clinical Validation, and Future Directions

Abstract

Skull fracture diagnosis is regarded as a significant diagnostic problem because it features a misdiagnosis rate up to 14.8% even among experienced radiologists. In this regard, the paper presents a systematic synthesis of 20 studies on skull fracture detection using conventional image-processing techniques, deep learning algorithms, and clinical validation. The best results obtained using the conventional technique are the recognition rates equal to 99-100%. In terms of deep learning models, their effectiveness is higher compared to others; for instance, CNN hybrid model reaches 97.6% accuracy, and YOLOv8 reaches 49% with recall 91.5%. Post-mortem skull fracture CT diagnosis features sensitivity 0.89 (0.80-0.94) with a base sensitivity of 0.87, and there are several visualization techniques such as azimuthal equidistant mapping that provide better skull fracture detection in less than one second of computational time. Important limitations are represented by the lack of dataset (CQ500 includes only 491 cases, including 84 patients with fractures), low inter-observer agreement (ICC < 0.4), and anatomical problems. Traumatic brain injury prevalence among maxillofacial fractures patients equals 51.04% (26,774).