Novel algorithm for knee localization and diagnosis and grading of knee osteoarthritis based on a priori information: data from OAI.
Authors
Affiliations (4)
Affiliations (4)
- Department of Orthopedics, Central Hospital of Shenyang Medical College, Shenyang, China.
- Platform Engineering Research Center, Neusoft Research Institute of Healthcare Technology, Shenyang, China.
- Department of Radiology, Affiliated Hospital of Liaoning University of Traditional Chinese Medicine, Shenyang, China.
- Department of Rehabilitation, Shengjing Hospital of China Medical University, No.16, Puhe Street, Shenyang North New Area, Shenyang, Liaoning Province, 110134, China. [email protected].
Abstract
To develop and evaluate a deep learning-based method for fully automatic segmentation of the knee joint and to diagnose and classify knee osteoarthritis (KOA ) using prior information. All X-rays were obtained from the Osteoarthritis Initiative (OAI). We comprehensively consider the a priori information of knee data, redesign the KOA assessment process, and propose the Anchor-free Knee Probability Calculation Net (AKPCNet) knee joint region of interest extraction algorithm by calculating the probability of each point as the left and right knee joint centroids, determining the location of the centroid of the knee joint, and then extracting the region of fixed size resolution around the centroid. We propose Attention Pooling, a global pooling optimization algorithm, and an attention pooling based low-order feature reinforcement network (APLFRNet) to improve the KL classification accuracy of KOA. The performance of the classification models was assessed using the area under the receiver operating characteristic curve (ROC AUC) and balancing accuracy. In total, 35,000 knee radiographs (anteroposterior view) were obtained from the Osteoarthritis Initiative (OAI). The accuracy of automatic recognition of left and right knee joint center points was 97.8% and 97.4%, respectively. The balancing accuracy of knee osteoarthritis assessment of five classifications according to the grading was 73.23%, the balancing accuracy of three classifications (KL0-1 vs.KL2vs.KL3-4) for estimating severity reached 82.22%, the accuracy rate of two classifications (KL0-1vs. KL2-4) for diagnosis was 87.6%. The accuracy of early diagnosis (KL0 vs. KL1 and KL0 vs. KL2) was 66.58% and 87.1%, respectively average areas under the curve of the five classifications, three classifications, and two classifications were 0.90,0.95 and 0.94 respectively. Our new algorithm can accurately assess the severity of knee osteoarthritis (KOA) and provide decision support for research and clinical practice.