Deep learning-assisted knee osteoarthritis automatic grading on plain radiographs: the value of multiview X-ray images and prior knowledge
Wei Li, Zhongli Xiao, Jin Liu, Jiaxin Feng, Dantian Zhu, Jianwei Liao, Wenjun Yu, Baoxin Qian, Xiaojun Chen, Yijie Fang, Shaolin Li
Abstract
Background: Knee osteoarthritis (OA) is harmful to people's health. Effective treatment depends on accurate diagnosis and grading. This study aimed to assess the performance of a deep learning (DL) algorithm based on plain radiographs in detecting knee OA and to investigate the effect of multiview images and prior knowledge on diagnostic performance. Methods: In total, 4,200 paired knee joint X-ray images from 1,846 patients (July 2017 to July 2020) were retrospectively analyzed. Kellgren-Lawrence (K-L) grading was used as the gold standard for knee OA evaluation by expert radiologists. The DL method was used to analyze the performance of anteroposterior and lateral plain radiographs combined with prior zonal segmentation to diagnose knee OA. Four groups of DL models were established according to whether they adopted multiview images and automatic zonal segmentation as the DL prior knowledge. Receiver operating curve analysis was used to assess the diagnostic performance of 4 different DL models. Results: The DL model with multiview images and prior knowledge obtained the best classification performance among the 4 DL models in the testing cohort, with a microaverage area under the receiver operating curve (AUC) and macroaverage AUC of 0.96 and 0.95, respectively. The overall accuracy of the DL model with multiview images and prior knowledge was 0.96 compared to 0.86 for an experienced radiologist. The combined use of anteroposterior and lateral images and prior zonal segmentation affected diagnostic performance. Conclusions: The DL model accurately detected and classified the K-L grading of knee OA. Additionally, multiview X-ray images and prior knowledge improved classification efficacy.