Purpose: Current bone mass screening guidelines are limited to women aged ≥65 and postmenopausal women with certain risk factors. Screening rates using Dual X-ray Absorptiometry (DXA) in this cohort without prior history of osteoporosis were 27% or less depending on age, with significant socioeconomic disparities in screening rates. We developed an opportunistic diagnosis algorithm to identify low bone mass in multi-vendor abdominopelvic CT exams using artificial intelligence (AI).
Methods: Matched abdominopelvic CT scans and DXA scans from 4406 patients aged ≥50 years were collected. The binary classification models were trained based on the T-scores reported in the scan to identify patients with low bone mass (T-score ≤-1). Image data was cleaned and normalized using conventional image processing and then input into a Densely Connected Convolutional Networks (Densenet121) architecture. Separate models were created using coronal and axial sections.
Results: Ten-fold cross validation was used to optimize hyperparameters for each model. An 80%-20% split was used to test the model. Mean area under the curve (AUC) for receiver operator characteristic curves was 0.83 for both models. Activation maps indicated that focus was on L3 vertebra in axials view and both hip and spine areas in coronal views.
Conclusion: A CT screening algorithm was developed to identify patients with low bone density consistent with osteopenia and osteoporosis. The high AUC indicates that this algorithm may be suitable to opportunistically screen patients undergoing abdominopelvic CT scanning.
Funding Support, Disclosures, and Conflict of Interest: Roubos Family Fund in Research
IM/TH- Image Analysis (Single Modality or Multi-Modality): Computer-aided decision support systems (detection, diagnosis, risk prediction, staging, treatment response assessment/monitoring, prognosis prediction)