Study Looks at Deep Radiomics Approach for Diagnosing Osteoporosis on Hip X-Rays

Deep radiomics models that included deep learning features had a 40 percent or greater increase in the specificity rate for diagnosing osteoporosis on hip radiographs in comparison to models that only emphasized clinical and/or textural features.

Emerging research suggests that a deep radiomics model, which incorporates clinical, textural, and deep learning features, may enhance the diagnosis of osteoporosis on hip radiographs.

In the retrospective study, recently published in Radiology: Artificial Intelligence, researchers assessed the use of seven deep radiomics models for diagnosing osteoporosis in 444 patients who had hip or pelvic AP radiographs and dual-energy X-ray absorptiometry (DXA).

The study authors found that the deep radiomics model that combined deep learning features with clinical and textural features (model-DTC) had an area under the curve (AUC) of .95, 89 percent sensitivity and the highest specificity at 83 percent. In contrast, the model that only emphasized clinical features (model-C) had higher sensitivity (99 percent) but only 31 percent specificity and a lower AUC (.83). The model that only emphasized textural features (model-T) had a lower AUC (.77) and sensitivity (77 percent) and a 55 percent sensitivity rate, according to the study.

“Our study results showed that deep radiomics models can be used to diagnose osteoporosis on hip radiographs with high diagnostic performance … and demonstrate feasibility as an opportunistic diagnostic tool,” wrote Hye Jin Yoo, M.D., Ph.D., who is affiliated with the Department of Radiology at the Seoul National University College of Medicine in Korea, and colleagues.

According to the study, 356 of the patients were women and the mean age of the study population was 70. In addition to model-DTC, model-C and model-T, Yoo and colleagues assessed model-D (emphasizing deep learning features), model-TC (textural and clinical features), model-DC (deep learning and textural features) and model-DT (deep learning and textural features).

For the four models that incorporated deep learning features, the authors noted an AUC range between 0.92-0.95 in contrast to an AUC range between 0.77 to 0.83 for the three models without deep learning features. In the models without deep learning features, Yoo and colleagues noted the specificity rate ranged between 31 to 55 percent in comparison to a range of 71 to 83 percent for the models with deep learning features.

Given the retrospective nature of the study, the authors acknowledged the possibility of patient selection bias. They noted that their model may be limited when it comes to treatment monitoring as they did not incorporate regression methods that could facilitate continuous reflection of bone mineral density. The study authors also pointed out they did not assess the model for its ability to predict the risk of fragility fracture.