A fully automated body analysis tool for abdominal CT can provide comprehensive prediction of cardiovascular risk, generating results nearly equivalent to those of manual segmentation in a fraction of the time.
Magudia
"The combination of automated body composition analysis of abdominal CT with our age-, race- and gender-specific reference curves have revealed a new biomarker, normalized visceral fat area, that can augment established cardiovascular risk models," said RSNA 2020 presenter Kirti Magudia, MD, PhD, a clinical fellow in radiology at the University of California, San Francisco School of Medicine.
As a radiology resident at Brigham and Women's Hospital in Boston, Dr. Magudia and her research team analyzed initial scans from all adult patients who underwent abdominal CT at Brigham and Women's, Massachusetts General Hospital and the Dana-Farber Cancer Institute in 2012. The study excluded patients with major cardiovascular disease or cancer. The cohort ultimately included 12,128 patients.
The team's analysis model begins with a full abdominal CT scan from PACS and applies selection logic to select only axial abdominal CT series.
"We then apply a slice selection network, which selects a slice at the level of the L3 vertebral body, and then a segmentation network is applied to produce a segmentation mask, which can be used to calculate body composition areas," Dr. Magudia explained.
"We found that the analysis showed high correlation and equivalency to manual segmentation," Dr. Magudia said. "Visceral fat area Z-scores derived from body composition reference curves are associated with subsequent myocardial infarction and subsequent stroke."
Body composition is largely thought of in terms of body mass index (BMI), which is associated with cardiovascular risk. However, there are some limitations to BMI, Dr. Magudia said. She displayed side-by-side abdominal CT images from two different patients with identical BMI but with a significantly different ratio of subcutaneous fat to skeletal muscle.
Calculations of the ratios of subcutaneous fat, visceral fat, and skeletal muscle on CT provide a more comprehensive picture of a patient's risk. Reference curves also incorporate data specific to age, race and sex.
"While manual calculation is the standard for risk analysis, it can be time-intensive and costly, and is usually limited to small cohorts and well-funded research studies," Dr. Magudia explained. "We demonstrate how automated analysis can be applied on a population scale amenable to clinical implementation and large-scale research."
The automated analysis was validated by comparing automated segmentation with radiologist segmentation, and the researchers found a very high correlation — 0.99 — for all body compartments.
"CT-based automated body composition analysis can unlock latent information from abdominal CT that will be useful to patients and physicians in clinical practice as well as in research," Dr. Magudia said.
"We demonstrate in this work that this analysis can be applied on a large scale with a low failure rate of 2.5%," Dr Magudia added. "With further research, this work could improve the care of a vast number of patients."
For More Information:
View the RSNA 2020 session Prediction of Major Cardiovascular Events in a Large Outpatient Adult Cohort Using Fully Automated and Normalized Deep Learning Body Composition Analysis of Routine Abdominal CT — SSGI04 at RSNA2020.RSNA.org.
