Automatic Measurement of Knee Joint Space Width on Anteroposterior X-ray Images Based on U2-Net

Deep learning can extract complex features from various data. To improve the accuracy and efficiency of clinical diagnosis and treatment, deep learning has been widely used in the processing of medical images. X-ray images can show typical radiological features such as narrowed joint space in patients with knee osteoarthritis (KOA), and the Kellgren-Lawrence (KL) grading system is commonly used to describe the severity and progression of the disease. However, the KL grading system has limited sensitivity in detecting early-stage disease changes. Therefore, measuring the joint space width (JSW) can more accurately assess whether the joint space has narrowed, which is an important means of evaluating osteoarthritis. Based on the experimental data of knee joint anterior-posterior X-ray images provided by the hospital, target detection is used to obtain image slices containing only the knee part. For the knee images, U2-Net is used to detect the knee joint space, and the contour of the entire joint space is obtained. The defined L1 norm is used to determine the effective range of the joint space, which is used to measure the maximum JSW and the average JSW. Comparing the experimental results with the actual joint space measurements determined under the guidance of professional doctors, the results show that the relative errors of all maximum JSW measurements were within 1.30%, and the relative errors of all average JSW measurements were within 3.90%, indicating that the automatic measurement model has good consistency with the doctor’s manual measurement.