Trainable rule-based algorithm for the measurement of joint space width indigital radiographic images of the knee

The progression of osteoarthritis (OA) can be monitored by measuring the mi nimum joint space width (mJSW) between the edges of the femoral condyle and the tibial plateau on radiographs of the knee. This is generally performed by a trained physician using a graduated magnifying lens and is prone to t he subjectivity and variation associated with observer measurement. We have developed software that performs this measurement automatically on digitiz ed radiographs. The test data consisted of 180 digitized radiographs of the knee (90 duplicate acquisitions) from 18 normal (nonarthritic) subjects an d 38 images from 10 subjects with OA. These were digitized and manually cro pped so that the images were free of nonanatomical structures and the knee was approximately centered. The software first determined the edge of the f emoral condyle on 400 mu m pixel subsampled images. Contours marking the lo cation of the tibial plateau in the medial compartment were found on 100 mu m images using the femoral edge as a reference. The algorithm was trained using an independent but similar data set and using a jackknife approach wi th the test data. The results were compared to contours drawn by a trained reader and the duplicate acquisitions were used to measure the reproducibil ity of the mJSW measurement. The reproducibility was 0.16 mm and 0.18 mm fo r normal and osteoarthritic knees, respectively, representing an improvemen t of approximately a factor of 2 over manual measurement. The algorithm als o showed excellent agreement with the hand-drawn contours and with mJSW det ermined by the manual method. (C) 2000 American Association of Physicists i n Medicine. [S0094-2405(00)01503-0].