Key-Point Detection Algorithm of Deep Learning Can Predict Lower Limb Alignment with Simple Knee Radiographs

Abstract

(1) Background: There have been many attempts to predict the weight-bearing line (WBL) ratio using simple knee radiographs. Using a convolutional neural network (CNN), we focused on predicting the WBL ratio quantitatively. (2) Methods: From March 2003 to December 2021, 2410 patients with 4790 knee AP radiographs were randomly selected using stratified random sampling. Our dataset was cropped by four points annotated by a specialist with a 10-pixel margin. The model predicted our interest points, which were both plateau points, i.e., starting WBL point and exit WBL point. The resulting value of the model was analyzed in two ways: pixel units and WBL error values. (3) Results: The mean accuracy (MA) was increased from around 0.5 using a 2-pixel unit to around 0.8 using 6 pixels in both the validation and the test sets. When the tibial plateau length was taken as 100%, the MA was increased from approximately 0.1, using 1%, to approximately 0.5, using 5% in both the validation and the test sets. (4) Conclusions: The DL-based key-point detection algorithm for predicting lower limb alignment through labeling using simple knee AP radiographs demonstrated comparable accuracy to that of the direct measurement using whole leg radiographs. Using this algorithm, the WBL ratio prediction with simple knee AP radiographs could be useful to diagnose lower limb alignment in osteoarthritis patients in primary care.

Keywords: convolutional neural network; knee; machine learning; prediction; weight-bearing line.

Figure 1

Pipeline of the study. Annotations: labeled image. Training set: the data sample used to fit the model. Validation set: the data sample used to provide an unbiased evaluation of a model fit on the training dataset while tuning the model hyperparameters. Test set: the data sample used to provide an unbiased evaluation of the final model fit on the training dataset.

Figure 2

Comparison of WBL mean accuracy in 4 and 27 points. (A) Comparison graph of 4 and 27 points in WBL error percentage. (B) Comparison graph of 4 and 27 points in pixel threshold. Validation set: the data sample used to provide an unbiased evaluation of a model fit on the training dataset while tuning the model hyperparameters. Test set: the data sample used to provide an unbiased evaluation of a final model fit on the training dataset.

Figure 3

Prediction and correct point. (A–D) Prediction point vs. correct point. (A) Prediction point in Rt knee, (B) correct point in Rt knee, (C) prediction point in Lt knee, (D) correct point in Lt knee. Prediction point: point predicted by the deep learning algorithm. Correct point: labeled point.

Figure 4

Histogram of the WBL ratio distribution in the validation and test sets. (A–D) Prediction distribution vs. target distribution. (A) WBL ratio prediction distribution in the validation set, (B) WBL ratio target distribution in the validation set, (C) WBL ratio prediction distribution in the test set, (D) WBL ratio target distribution in the test set. WBL ratio prediction: predicted value of the WBL, WBL ratio target: calculated WBL target.

Figure 5

Graph of the WBL mean accuracy. (A) Accuracy graph of the WBL error percentage value, (B) accuracy graph of the pixel threshold. Validation set: the data sample used to provide an unbiased evaluation of a model fit on the training dataset while tuning the model hyperparameters. Test set: the sample of data used to provide an unbiased evaluation of a final model fit on the training dataset.