Assessment of Osteoporotic Fracture Risk in Lumbosacral Spine Radiographs using Convolutional Neural Network
- Nov. 2020
- by Jaewon Lee et. al.
To develop and evaluate a deep learning-based osteoporotic fracture risk prediction model in lumbosacral spine radiographs.
METHOD AND MATERIALS
Lumbosacral spine radiographs were retrospectively collected as baseline and follow-up pairs from 491 men and 1,333 women with ages ranging from 50 to 75 years showing no fracture at baseline study. In the follow-up studies, 293 cases of compression fracture were reported. The average period between baseline and follow-up was 40.71±26.84 months.
A convolutional neural network (CNN) was trained to predict compression fracture occurrence in follow-up studies given baseline images. A key-point detection network was trained to extract region of interest (ROI) patches corresponding to L1-L5 vertebral bodies, and the extracted ROIs were aligned vertically and horizontally using geometric transformations. Another CNN was then trained to predict osteoporotic fractures from ROI patches, and area under the receiver operating characteristic curves (AUCs) were evaluated using 5-fold cross-validation. A naive-Bayes model served as a baseline, predicting osteoporotic fracture risk using only the subject’s age.
Overall the CNN and naive-Bayes models were able to predict osteoporotic fracture risk at AUCs of 0.75±0.02 and 0.69±0.05 respectively. Further, the two models achieved AUCs of 0.78±0.02 and 0.72±0.02 for cases with follow-up period exceeding 5 years (60 months), and 0.70±0.02 and 0.69±0.05 for cases with less than 5 years. When subjects were split by gender, the AUCs achieved were 0.77±0.02 and 0.68±0.01 for women and 0.71±0.04 and 0.67±0.11 for men, in order of CNN and naive-Bayes.
A CNN outperformed an age-based model in predicting osteoporotic fracture in lumbosacral spine radiographs.
Deep learning can potentially be used as a CAD system to predict osteoporotic fracture risk in lumbosacral spine radiographs.