I'm going from hip joint to the ankle.
A close-up knee image cannot fully show whole-leg loading and bone angles. Standing full-length X-rays include the hip, knee and ankle. Planning arthroplasty or osteotomy involves placing landmarks and connecting lines to measure alignment; we developed a way to automate this repeated work.
First, you find the joint and then you find a point inside it.
11,212 images were used to develop and evaluate the model. After locating hip, knee and ankle regions, it identifies 15 anatomical landmarks—the starting points for angle measurement. Specialist reference measurements were used; 305 images not used for training formed a separate test set.
So I repeated the same angle, and I checked to see if it could be solved.
Agreement between the model and reference measurements was high for each angle, and repeated measurement of the same image was consistent. Mean absolute errors were below 0.52° for angles other than JLCA. A small average error does not mean that every image has a small error. Outliers and angle-specific performance should be checked together.
joint line read the angle between them carefully.
JLCA is the small angle between the femoral and tibial joint surfaces. It should not be described as having the same accuracy as the other larger alignment angles. Degenerative blurring of the boundaries and differences in positioning rotation may change landmark detection. The original paper also shows difficult images so readers can inspect what the model detected.
Automatic measurement is followed by the doctor's interpretation.
Fast angle measurement and choosing a surgical target are different tasks. Clinicians still assess bony versus intra-articular deformity, symptoms, and movement. This study was internal validation at one institution, so confirmation elsewhere is needed. Repeatable measurements are a foundation for better research and planning.