E3315. DeepMetabolics: An Overview of Artificial Intelligence Assisted Tools to Automatically Assess Metabolic Syndrome on Imaging
Authors
Abhinav Suri;
National Institute of Health Clinical Center
Pritam Mukherjee;
National Institute of Health Clinical Center
Ronald Summers;
National Institute of Health Clinical Center
Background
Metabolic syndrome affects millions of adults each year in the United States alone. It puts individuals at increased risk of adverse cardiovascular events such as stroke, heart attack, and more. Moreover, tests to identify metabolic syndrome such as lipid panels and A1c often rise after metabolic syndrome has taken hold. Additionally, body weight metrics such as BMI and waist circumference have also been found to be flawed in populations. As a result, patients with subtle signs of metabolic syndrome go undiagnosed for years.
Opportunistic identification of metabolic disease on imaging offers another opportunity to “catch” cases of metabolic syndrome as they develop, leading to increased diagnosis and a larger window of opportunity for intervention.
Educational Goals / Teaching Points
Teach learners about definitions of metabolic syndrome, discuss current hematological screening guidelines and pitfalls, describe and depict imaging correlates of metabolic syndrome effects such pancreatic anatomical changes, coronary artery calcification, and muscle fat content on CT and MRI, describe metrics that quantify metabolic disease on imaging, and highlight tools that utilize computer vision techniques and deep learning that can aid in radiologists’ assessment of metabolic disease.
Key Anatomic/Physiologic Issues and Imaging Findings/Techniques
We aim to highlight assessment tools of metabolic syndrome in our presentation and give an overview of automated tools that can detect signs of metabolic syndrome on imaging. Deep learning networks offer the ability to detect diabetic vs nondiabetic pancreases on CT imaging. T2DM has also been identified using deep learning from frontal chest radiographs. Additionally, deep learning networks have been shown to predict the incidence of diabetes 7 years before formal diagnosis. Additionally, further studies have found that measurement of markers such as total adipose tissue, visceral adipose tissue, subcutaneous adipose tissue, skeletal muscle index, liver attention, and abdominal aortic Agatston score can also be automated.
Conclusion
Metabolic syndrome is a highly prevalent disease with many undiagnosed cases. Opportunistic screening via imaging offers an additional venue to detect cases ahead of time. There are many extant markers of metabolic syndrome and these markers have been successfully measured automatically, reducing time and effort on the part of reading radiologists.
