Wednesday, November 29 | 10:00 a.m.-10:10 a.m. | W3-SSMK08-4 | Room E450A
A deep learning-based framework for automated screening of osteoporosis on lumbar spine plain radiographs shows potential as another way to opportunistically make use of imaging studies performed for other indications, according to this presentation.
In a retrospective study led by presenter Bin Zhang, MD, of The First Affiliated Hospital of Jinan University in China, 1,538 pairs of lumbar spine radiographs and dual-energy x-ray absorptiometry (DEXA) results from females aged ≥ 50 years were collected between January 1, 2014, and October 10, 2019. Patients were categorized into osteoporosis and non-osteoporosis groups. The study included 1,147 patients in the training set, 145 patients in the validation set, 143 patients in the internal test set, and 103 patients in the external test set.
The fully automated framework involved a dual-view, two-stage, deep learning-based method; a U-net network (Model 1) performs automatic vertebral segmentation, and a ResNet18 (Model 2) model then detects osteoporosis.
Model 1 yielded average dice coefficients of 0.836, 0.837, and 0.838 for posterior-anterior view, and 0.858, 0.854, and 0.8 for lateral view on the validation, internal test, and external test sets, respectively. Model 2 based on ensemble views achieved an area under the curve of 0.869, 0.798, 0.795, and 0.763 in the training, validation, internal test, and external test sets, respectively. Zhang will discuss how the team arrived at these findings Wednesday.
"While the tool cannot replace [DEXA] for osteoporosis screening, it can be a valuable option when lumbar spine radiography is readily available, and [DEXA] has not been performed," the authors wrote. "Therefore, the clinical implementation of such a tool could enhance disease recognition and aid in preventing osteoporotic fractures."
