Automatic measurement of Caton index on knee X-ray images using a key point detection model

Abstract

Purpose

To explore the feasibility of the You Only Look Once (YOLO) algorithm in the measurement of Carton index.

Methods

1156 knee X-ray images were collected from two centers (960 and 196). Five key points at patella and tibia on knee X-ray were labeled using the software of Labelme. YOLO11 pose models (including YOLO11n, YOLO11m and YOLO11x) were refined by labeled images from center A, and was then used to detect keypoints on images from center B. A line was the line between anterior edge of the tibial plateau and the lower pole of patellar articular surface, and B line was patellar articular surface. Carton index (A/B ratio) of 196 cases was obtained by senior radiologist, junior radiologist and YOLO respectively. The Bland Altman plot, Pearson Correlation test, Mean Absolute Error (MAE) and Intra-class correlation coefficient (ICC) were used to evaluate the agreement in measurement.

Results

Carton index of 196 images were automatically obtained with YOLO11n-pose, YOLO11m-pose and YOLO11x-pose. The ICC between senior and junior radiologists was 0.89. Pearson correlation coefficients were 0.23, 0.43 and 0.73 respectively for YOLO11n, YOLO11m and YOLO11x. ICC were 0.23, 0.42 and 0.72 respectively for YOLO11n, YOLO11m and YOLO11x. MAE were 0.20, 0.17 and 0.10 respectively for YOLO11n, YOLO11m and YOLO11x.

Conclusions

YOLO11x-pose model shows promise in the automatic measurement of Carton index on the knee X-ray image.