A vendor-agnostic AI system using optical character recognition automatically drafts DEXA bone density reports, reducing report creation time by up to 60% and turnaround time by up to 75% while maintaining 99.9% numerical accuracy across academic and community imaging centers.
- AI system reduced median report turnaround time from 2.40 hours to 0.96 hours at academic sites and from 133.82 hours to 42.10 hours at community sites
- Report creation time dropped from 3.48 minutes to 0.87 minutes at academic sites and from 1.42 minutes to 0.63 minutes at community sites
- Numerical accuracy reached 99.9% for AI-generated drafts, exceeding the 99.4% accuracy of manual reports
- AI drafts achieved 100% completeness at community sites compared to 45% for original manual reports
- System works across multiple scanner types without requiring vendor-specific data feeds, enabling uniform deployment across healthcare settings
A vendor-agnostic AI system that automatically drafts dual-energy x-ray absorptiometry (DEXA) reports significantly reduces report creation time and report turnaround time, according to a study published August 13 in the Journal of the American College of Radiology.
A team at Thomas Jefferson University Hospital in Philadelphia deployed the system across two academic and two community outpatient imaging practices and evaluated its performance across multiple scanner types, with the system offering a potential alternative to repetitive, time-consuming, and error prone manual reporting, according to the group.
“The observed reductions in report editing time and turnaround time may be relevant in the context of ongoing radiologist workforce constraints and increasing imaging volumes,” noted first author Paras Lakhani, MD, and colleagues.
DEXA is the reference standard for assessing bone mineral density and estimating fracture risk. Despite its clinical importance, DEXA reporting is burdensome, the group wrote. Radiologists must accurately manually transcribe multiple numeric values such as bone mineral density, T-scores, and Z-scores across several anatomic regions while assigning diagnostic categories and generating longitudinal comparison statements.
Prior automation efforts have relied on modality-driven structured data feeds, which are typically vendor-specific and difficult to deploy uniformly across sites with heterogeneous scanner fleets, the researchers added. Alternatively, the group developed a vendor-agnostic system that uses AI-based optical character recognition (OCR) to extract measurements directly from standardized DEXA DICOM images and rule-based logic to generate complete draft reports within the radiology reporting system.
To test it, the researchers launched the system at two academic outpatient imaging sites in February 2025 and at two community sites in August 2025. They assessed its operational impact using a pre/post design, measuring report creation time (report start to first signature) and report turnaround time (study completion to final signature). Report accuracy was assessed by comparing AI drafts and original reports against source images (n = 100 per site).
According to the findings, at the academic sites, median report turnaround time fell from 2.40 to 0.96 hours (p < 0.0001), and median report creation time dropped from 3.48 to 0.87 minutes (p < 0.0001). Reductions were larger at the community sites, where median turnaround time decreased from 133.82 to 42.10 hours (an absolute reduction of nearly four days) and median report creation time fell from 1.42 to 0.63 minutes (both p < 0.0001).
In addition, at the academic sites, numerical accuracy was 99.9% for AI drafts versus 99.4% for originals (p = 0.022), with equivalent completeness and diagnostic accuracy. The gap was most pronounced at the community sites, where AI drafts achieved 100% completeness compared with 45% for original reports (p < 0.001), while maintaining equivalent numerical and diagnostic accuracy, the researchers reported.
“Implementation of a vendor-agnostic, AI-OCR–based DXA reporting system was associated with significant reductions in report creation time and report [turnaround time], while maintaining high numerical accuracy, diagnostic accuracy, and report completeness across both academic and community practice settings,” the group wrote.
Ultimately, the findings suggest that an image-based extraction approach can be integrated into routine clinical workflows without degradation of reporting quality, the researchers concluded.
The full study is available here.
