Automatic surgical classification of knee X-ray images using machine deep learning
10.3760/cma.j.cn115530-20240716-00297
- VernacularTitle:深度学习膝X线片实现手术类型的自动识别
- Author:
Qianli MA
1
;
Ming ZHENG
;
Qiang CHEN
;
Yuyun ZHENG
;
Jiongjiong GUO
;
Yumin CHEN
;
Yi ZHAO
Author Information
1. 福州市第二总医院骨科，福州 350007
- Keywords:
Tibial fracture;
Fracture fixation;
Arthroplasty, replacement, knee;
Osteotomy;
Deep learning
- From:
Chinese Journal of Orthopaedic Trauma
2024;26(10):834-841
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To evaluate the performance of our self-developed deep learning (DL) model which was designed to automatically classify the knee X-ray images into one non-surgical category and 4 surgical categories, including non-surgical knees (NSK), high tibial osteotomy (HTO), total knee arthroplasty (TKA), unicompartmental knee arthroplasty (UKA), and tibial plateau fracture fixation with an internal fixation plate (TPFF).Methods:The knee X-ray images were collected of the patients who had undergone knee joint surgery at Department of Orthopaedics, Fuzhou Second General Hospital from January 2017 to December 2022. On the Baidu EasyDL AI platform, a multi-class object recognition DL model was built using the You Only Look Once (YOLO) algorithm. The model was trained on a dataset of 1,281 knee anteroposterior X-ray images (including NSK, HTO, TKA, UKA, and TPFF) to generate a DL model which was able to automatically recognize and classify the knee X-ray images. The reliability of the model classification performance was evaluated by analyzing the 5 indicators [accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV)] using a test set of 450 knee anteroposterior X-ray images with the above 5 categories. The receiver operating characteristic curve was plotted and the area under the curve (AUC) was calculated to further quantify the classification performance of the model.Results:In the test set, on the whole, the model achieved an accuracy of 97.0%, a sensitivity of 92.4%, a specificity of 98.1%, a PPV of 92.4%, and a NPV of 98.1%, an AUC of 0.947, indicating a high reliability in classifying various categories. The model showed a best performance for TKA, with the 5 indicators being 99.1%, 99.0%, 99.1%, 97.1% and 99.7%. The model showed a slightly lower sensitivity for TPFF and HTO (87.0% and 86.0%, respectively).Conclusion:A successful DL model has been developed which can automatically classify the knee X-ray images into non-surgical and surgical categories due to its satisfactory performance, particularly in accuracy and AUC.
