A YOLOv8 deep learning model for detecting labral injury via hip magnetic resonance imaging.
Authors
Affiliations (4)
Affiliations (4)
- First Clinical Medical College, Gansu University of Chinese Medicine, Lanzhou, China.
- Department of Radiology, Nanbu People's Hospital, Nanbu, China.
- School of Nursing, Gansu University of Chinese Medicine, Lanzhou, China.
- Department of Radiology, Gansu Provincial Hospital of TCM, Lanzhou, China.
Abstract
Acetabular labral injury is a common traumatic lesion of the hip joint and one of the most common hip disorders that causes hip pain. Accuracy in the diagnosis of labral injury depends on the level of experience of the interpreting radiologist. This study aimed to determine the feasibility of using the You Only Look Once version 8 (YOLOv8) deep learning model for acetabular labral injury detection on proton density-weighted fast spin-echo (PD-FSE) magnetic resonance imaging (MRI) sequences. Adult patients who underwent oblique sagittal and oblique coronal plane MRI examinations between 2021 and 2025 were enrolled. PD-FSE sequence images were annotated to construct a dataset, and the YOLOv8 algorithm was used to develop the two-dimensional (2D) slice-level deep learning model. A total of 936 PD-FSE sequence images were collected from 111 patients. The dataset was split at the patient level into training (76 patients and 675 images), validation (23 patients and 169 images), and test (12 patients and 92 images) sets. A labrum-absent test set consisting of 58 images was established to evaluate the generalization performance of the models. The accuracy, sensitivity, and specificity of the optimal YOLOv8 model (YOLOv8m) were 0.88 [95% confidence interval (CI): 0.82-0.92], 1.00 (95% CI: 0.96-1.00), and 0.74 (95% CI: 0.62-0.82), respectively. No statistically significant difference was detected between the YOLOv8m model and senior chief radiologists (P=0.864), while the model significantly outperformed junior radiologists (P=0.017). The misjudgment rate of the optimal YOLOv8m model in the labrum-absent test set was 12.1%. This is the first study to establish a YOLOv8-based 2D slice-level model for detecting labral injury detection. The deep learning model constructed based on YOLOv8 demonstrated diagnostic performance comparable to that of senior radiologists in detecting acetabular labral injuries. This model demonstrates promising potential for future clinical decision support in suspected acetabular labral injury management.