RSNA 2016

Abstract Archives of the RSNA, 2016


IN212-SD-MOA3

Preliminary Results Using Deep Learning Artificial Intelligence to Estimate Bone Mineral Density on Abdominal CT Exams for Screening Osteoporosis

Monday, Nov. 28 12:15PM - 12:45PM Room: IN Community, Learning Center Station #3



Michael Y. Park, MD, Seoul, Korea, Republic Of (Presenter) Nothing to Disclose
Seung Eun Jung, MD, Seoul, Korea, Republic Of (Abstract Co-Author) Nothing to Disclose
Sangki Kim, PhD, Seoul, Korea, Republic Of (Abstract Co-Author) Nothing to Disclose
Moon Hyung Choi, MD, Seoul, Korea, Republic Of (Abstract Co-Author) Nothing to Disclose
Joon-Il Choi, Seoul, Korea, Republic Of (Abstract Co-Author) Nothing to Disclose
Soon Nam Oh, MD, Seoul, Korea, Republic Of (Abstract Co-Author) Nothing to Disclose
Sung Eun Rha, MD, Seoul, Korea, Republic Of (Abstract Co-Author) Nothing to Disclose
Jae Young Byun, MD, Seoul, Korea, Republic Of (Abstract Co-Author) Nothing to Disclose
PURPOSE

To determine whether deep learning artificial intelligence can be used to automatically estimate and screen for osteoporosis on abdominal CT exams.

METHOD AND MATERIALS

A total 4,203 pairs of lumbar spine dual-energy X-ray absorptiometry (DXA) exams and abdominal CT exams taken within 60 days of the same patient were included in this study. The median age of patients was 54 (IQR: 48-60), and the 3,973 (94.5%) of the cases had the DXA and CT exams on the same day (mean 0.4 days). The L1-L4 mean bone mineral density (BMD) values were used as a reference standard, and 3,783 (90%) randomly selected cases were used for deep learning artificial intelligence training using a convolutional neural network model (VunoNet, Seoul, Korea). The other 420 (10%) of cases were excluded from training and used to evaluate the deep learning results. Cutoff values for using deep learning estimated BMD to screen osteoporosis were analyzed. Osteoporosis for determining cutoff values in this study was defined as cases with BMD T-scores of less than or equal to -2.5, with T-scores calculated only using lumbar spine DXA BMD results using reference population BMD T-score tables without compensating for weight.

RESULTS

The median DXA BMD and median deep learning estimated BMD were 1.159 (IQR: 1.042-1.271) and 1.171 (IQR: 1.055-1.286). A high linear correlation (r2 = 0.714) between deep learning estimated BMD values and DXA BMD values were noted. When using a deep learning estimated BMD cutoff value of less than or equal to 0.929 g/cm2 for screening osteoporosis, a sensitivity of 94.4% and specificity of 97.0% could be achieved when using lumbar spine DXA results as a gold standard, with a high AUC of 0.990.

CONCLUSION

Deep learning artificial intelligence for estimating BMD on abdominal CT exams has the potential to be used to automatically screen for osteoporosis in routine abdominal CT exams.

CLINICAL RELEVANCE/APPLICATION

This study shows the potential for using deep learning artificial intelligence to fully automate screening for osteoporosis in routine abdominal CT exams taken for other reasons.