Anatomical landmark localization via convolutional neural networks for limb-length discrepancy measurements

摘要

Background Measurement of limb-length discrepancy (LLD) from a radiograph is a cognitively simple but time-consuming task. Objective To develop a convolutional neural network (CNN) to localize anatomical landmarks from full-length lower-extremity radiographs in predicting LLD. Materials and methods The author searched a hospital's image database to identify studies performed between Feb. 1, 2016, and Sept. 30, 2019. Inclusion criteria were: (1) patients <= 21 years old, (2) study indication of LLD, (3) full-length lower-extremity anteroposterior radiographs performed on the EOS system, and (4) imaging field-of-view that included entire bilateral femurs and tibias. The six requisite ground truth anatomical landmarks for measuring LLD from each radiograph - bilateral top of femoral heads, medial femoral condyles, and center of tibial plafonds - were manually labeled by a pediatric radiologist. For each landmark, a two-dimensional heatmap was generated to encode the pseudo-probability of a landmark being at a particular spatial location. A CNN was developed that regressed across a collection of these heatmaps for landmark localization and bone length predictions. Results The study cohort consisted of 504 full-length lower-extremity radiographs from 359 patients with wide ranging skeletal deformities and in situ hardware. Evaluation of this CNN showed that the mean point-error for the predicted top of femoral head, medial femoral condyle, and center of tibial plafond were 0.37 cm, 0.39 cm and 0.42 cm, respectively. The mean absolute error for the predicted femoral, tibial and limb lengths, and LLD were 0.33 cm, 0.30 cm, 0.30 cm, and 0.36 cm, respectively. Predicted bone lengths correlated strongly with ground truth. Conclusion This prototype CNN delivered promising results in predicting bone lengths from full-length lower-extremity radiographs and offers the potential use of a computer algorithm to predict LLD.