YOLOv12 Algorithm-Aided Detection and Classification of Lateral Malleolar Avulsion Fracture and Subfibular Ossicle Based on CT Images: Multicenter Study

Comprehensive Summary

In this study, Liu et al. evaluated the ability of Deep Convolutional Neural Networks (DCNNs) to detect and classify Lateral Malleolar Avulsion Fractures (LMAFs) and Subfibular Ossicles (SFOs) in CT images. The dataset was compiled from a cohort of 1918 patients, with images being split into a training set, an internal validation set, and an external validation set. The initial distinction between LMAF and SFO was made by a senior radiologist based on MRI findings. 4 DCNNs were evaluated: YOLOv12, faster R-CNN, SSD, and RetinaNet. Radiologists were also asked to evaluate each case based on CT images, and clinician performance was compared to that of AI. The mean Average Precision (mAP) scores for each of DCNN were also evaluated. YOLOv12 had the strongest performance (mAP = 92.1%), while the other 3 models showed significant room for improvement (R-CNN 63.7%, SSD 63.0%, RetinaNet 67.0%). YOLOv12 also had high AUC, accuracy, and sensitivity scores (0.983, 0.920, and 0.867 respectively), all of which outperformed the radiologists (AUC= 0.755, Accuracy= 0.756, Specificity= 0.750). Overall, the YOLOv12 model shows promise in detecting LMAFs and SFOs in a clinical setting. Future direction should work toward increasing its robustness and applicability in large-scale settings.

Outcomes and Implications

Due to the similar characteristics of LMAFs and SFOs, it is often difficult to differentiate between them using CT imaging. The use of a DCNN such as YOLOv12 to aid clinicians has several important benefits. First, by improving the accuracy and timeliness of detection, these tools can help guide more efficient treatment planning and ultimately improve patient outcomes. Additionally, workflow efficiency can be enhanced, especially in high volume settings where radiologists will be able to reduce reading time. As the use of DCNNs in fracture detection continues to grow, these tools have the potential to become valuable components to routine clinical workflows and improve overall patient care.