Determining the Mechanical Axis of the Femur From a Standard Antero-Posterior Knee Radiograph With Deep Learning

As the femoral head is not visible on a standard antero-posterior (AP) knee view, surgeons looking to assess alignment on a patient who has only an AP view must either (i) forgo crucial long leg measurements, (ii) use the less meaningful anatomic axis, or (iii) use crude approximations. The purpose of this study was to develop a deep learning (DL) model for predicting the correction factor between the anatomical and mechanical axes.

We queried our institutional image registry for patients who had both a bilateral AP knee and a long-leg radiograph taken within 90 days of each other. We used two validated DL algorithms to measure the mechanical axis of each femur on the long leg view and the corresponding anatomic axis from the AP view. Then, we trained a DL regression model to predict the correction factor based on the AP image alone. We compared the mean absolute error (MAE) and standard deviation of MAE (SD) of our DL model against two other published approaches (i) linear regression and (ii) adding 6° of varus to the anatomic axis.

The DL approach to predicting the correction factor between anatomic and mechanical axes resulted in an MAE of 1.02° (SD: 0.94°). The linear regression approach resulted in an MAE of 1.34° (SD: 1.10°), while the 6° varus approach resulted in an MAE of 2.00° (SD: 1.47°).

We developed a DL model to predict the mechanical axis of the femur from a single, standard AP view knee radiograph. The DL model was accurate to within 1°, a clinically relevant level of accuracy, and a substantial improvement on published regression and 6° varus addition methods. This model delivers automated calculations of several important alignment measurements from standard radiographs, which may provide substantial clinical and research benefits.