Yolov7 neural network for pelvic X-rays after total hip arthroplasty: automatically measuring component position parameters- a retrospective feasibility study.
Authors
Affiliations (8)
Affiliations (8)
- Department of Imaging, The Second Affiliated Hospital of Baotou Medical College, Baotou, 014030, China.
- Department of Orthopedics, The Second Affiliated Hospital of Inner Mongolia Medical University, Hohhot, 010000, China.
- Department of Orthopedics, The Second Hospital of Ulanqab, Ulanqab, 011800, China.
- Department of Imaging, The Second Affiliated Hospital of Inner Mongolia Medical University, Hohhot, 010000, China.
- Department of Human Anatomy, Inner Mongolia Medical University, Xinhua Street, Hohhot, 010000, China.
- Department of Imaging, the Affiliated Hospital of Inner Mongolia Medical University, Hohhot, 010000, China.
- Department of Imaging, Baotou Cancer Hospital, Baotou, 014030, China. [email protected].
- Department of Human Anatomy, Inner Mongolia Medical University, Xinhua Street, Hohhot, 010000, China. [email protected].
Abstract
After total hip arthroplasty (THA), regular follow-up X-rays are required to evaluate the position of the prosthesis. However, measuring the parameters of prosthesis position consumes a significant amount of time, and there may be variations in the results obtained by different doctors. To validate the feasibility of an artificial intelligence model for the automated measurement of component position parameters on anteroposterior X-rays after THA. Post-THA anteroposterior X-rays were collected from five hospitals in Inner Mongolia, China. They were divided into training, validation and test sets. The model was divided into left, right, and bilateral THA modules for training. Statistical analysis was performed using Python 3.8 and SPSS 25.0. The accuracy of the model in keypoint detection and measurement were assessed. In addition, we compared the time difference between the model and manual measurements. A total of 1050 X-rays were included in the training and validation sets (350 X-rays for each module), and 105 in the test set (35 X-rays for each module). Overall, 80.00-100.00% of the model-predicted keypoints fell within 3 mm of the standard keypoints. The model-predicted values were highly consistent with the standard reference values (intra-class correlation coefficient = 0.89-0.99, r = 0.83-0.99, root mean square error = 0.96-3.43, mean absolute error = 0.78-2.87, mean difference = -1.29-1.17). The model-predicted values were highly consistent with the measurements of senior attending radiologists. The model took an average of 3.2 s to measure an X-ray image, which is significantly shorter than the 751.76 s in manual measurement. This study validates the feasibility of the model for rapid and accurate keypoint identification and component position parameter measurement on anteroposterior X-rays. The model holds promise as a diagnostic aid to alleviate clinicians' workload and facilitate the monitoring of prosthetic positioning changes following THA. Not applicable.