A deep learning approach to automatically quantify lower extremity alignment in children.

Abstract

To develop and validate a convolutional neural network (CNN) capable of predicting the anatomical landmarks used to calculate the hip-knee-ankle angles (HKAAs) from radiographs and thereby quantify lower extremity alignments in children. A search of the image archive at a large children's hospital was conducted to identify full-length lower extremity radiographs performed in children (≤ 18years old) for the indication of lower extremity alignment (7/2019-10/2019). A radiologist manually labeled each radiograph's six requisite anatomical landmarks used to measure HKAAs (bilateral centers of the femoral head, tibial spine, and tibial plafond) and defined the resultant labels as ground truth. A 2D heatmap was generated for each ground truth landmark to encode the pseudo-probability of a landmark being at a particular location. A CNN was developed for indirect landmark localization by regressing across a collection of these heatmaps. The landmarks predicted from this model were used to calculate the HKAAs. Absolute prediction error and intraclass correlation were used to assess the accuracy of the HKAA estimates. The study cohort consisted of 528 radiographs from 517 patients (mean age = 10.8years, SD = 4.2years). Evaluation of this CNN showed few HKAA prediction outliers (12/1056 [1.1%]), defined as having an absolute prediction error of > 10°. Excluding these outliers, the study cohort's mean absolute prediction error for the HKAA was 0.94° ± 0.84°, and the intraclass correlation between the ground truth and prediction was 0.974. The proposed CNN generated promising results and offers potential for using this model as a computer-aided diagnostic tool.