Augmenting generalizability of automated methods with local fine-tuning can help providers diagnosis osteoporosis with these scans.
Identifying patients who have osteoporosis can be easier on abdominal CT scans with automated methods that have improved generalizability.
It is well known that measuring the Hounsfield unit (HU) in the L1 vertebral body (VB) is an opportunistic way to screen patients for osteoporosis. But, the HU thresholds may not generalize well between different groups based on population and imaging protocol differences.
To see whether it is possible to augment opportunistic osteoporosis screening, Steven Rothenberg, M.D., assistant professor of diagnostic radiology at the University of Maryland School of Medicine, compared two deep learning model approaches to an external institution. His team first applied previously established thresholds to automatically extracted L1 HU, and they also trained a locally fine-tuned model using automatically extracted HUs, bone mineral densities (BMD), and demographic variables.
Rothenberg presented the team’s results in a poster during the Society for Imaging Informatics in Medicine (SIIM) 2021 Virtual Annual Meeting.
For more SIIM 2021 coverage, click here.
For the study, the team used a deep learning model that had been trained on images from one institution to automatically segment lumbar VBs on abdominal CTs and to extract trabecular HUs from each body in order to assess the external dataset from their local site.
They gathered 200 consecutive abdominal CTs from their hospital that had undergone DEXA examination within six months over a two-year timeframe. They, first, applied L1 VB HU cut-offs to diagnose potential osteoporosis, and, then, they used age and sex data alongside automatically-extracted L1 and L2 HU and BMD to train both decision tree and random forest models for osteoporosis diagnosis.
Based on their analysis, the locally fine-tuned random forest and decision tree models had average F1 scores of 0.84 and 0.8, respectively. These results outperformed the 0.54 and 0.46 scores achieved by L1 HU thresholds of 80 and 120, respectively. In addition, they team determined the average sensitivity for the fine-tuned models were greater than 80 percent – far outpacing the 40 percent and 57 percent achieved by the two HU thresholds.
Overall, the team said, their results indicate that local models that are fine-tuned on demographic variables and automatically-extracted VB HU and BMD data do result in improved generalization for opportunistic osteoporosis screening.
For more coverage based on industry expert insights and research, subscribe to the Diagnostic Imaging e-Newsletter here.
Can Emerging AI Software Offer Detection of CAD on CCTA on Par with Radiologists?
May 14th 2025In a study involving over 1,000 patients who had coronary computed tomography angiography (CCTA) exams, AI software demonstrated a 90 percent AUC for assessments of cases > CAD-RADS 3 and 4A and had a 98 percent NPV for obstructive coronary artery disease.
Mammography Study Compares False Positives Between AI and Radiologists in DBT Screening
May 8th 2025For DBT breast cancer screening, 47 percent of radiologist-only flagged false positives involved mass presentations whereas 40 percent of AI-only flagged false positive cases involved benign calcifications, according to research presented at the recent American Roentgen Ray Society (ARRS) conference.
MRI Study at ARRS Raises Questions About Disparities in Detection of MASLD
May 3rd 2025New research revealed that Hispanic Americans with evidence of hepatic steatosis on MRI but no formal diagnosis of MASLD had over a fourfold higher risk of developing hepatocellular carcinoma in comparison to those who had a formal diagnosis of MASLD.