Deep learning for indirect and direct cardiovascular risk estimation from ungated chest CT
Accurately targeting medical therapies is central to management of patients with coronary artery disease (CAD), regardless of the presence of symptoms. Coronary artery calcium (CAC) scoring from ECG-gated, non-contrast, computed tomography (CT) scans has emerged as a leading biomarker of CAD and cardiovascular risk. We have demonstrated that deep learning (DL) of CAC scores from low dose ungated chest CT from PET/CT and SPECT/CT can accurately predict adverse cardiovascular events with analysis performed in < 6 seconds per patient, making it suitable for large scale batch processing. However, there is also an abundance of information other than CAC contained within a chest CT, such as cardiac volumes and epicardial adipose tissue, which could be used to provide risk stratification for patients. Therefore, it may be possible to further refine risk stratification from ungated chest CT beyond that possible with CAC alone. It could be especially valuable to explore this in a large dataset like the National Lung Screening trial (NLST).
We propose to develop and validate state-of-the art DL for cardiovascular risk estimation from the low dose chest CT. We propose these 3 specific aims.
1) Derive standard measures of cardiovascular risk with DL. We will leverage and adapt our existing DL models to segment and quantify CAC, epicardial adipose tissue volume and attenuation, cardiac volumes, and liver to spleen attenuation ratios from low-dose chest CT. This would allow risk predictions to consider CAC information and cardiac volumes, but also other known features of importance present of chest CT.
2) We will determine whether these measures can be efficiently combined to predict cardiovascular risk. We will combine CAC and associated information such as density, number of lesions, and cardiac volumes, with non-cardiac measures such as liver attenuation, spleen attenuation, liver/ to spleen attenuation ratio,and skeletal muscle volume to provide risk stratification for all-cause mortality and cardiovascular mortality in patients from the NLST, without exclusions and specifically assessing sub-groups of patients (sex, race, previous CAD). Integration will be accomplished by extreme boosting models as in our previous work.
3) We will explore direct cardiovascular risk prediction from CT image data. This approach can allow for identification of latent DL image features (without specific segmentation of structures), which are associated with risk of cardiovascular events. Attention maps will be integrated into the model to highlight image regions associated with increased risk, allowing physicians to review the validity of the model and understand the findings as well as provide insights into novel imaging biomarkers.
The large set of chest CT data from NLST (n=22417) has already been transferred to Cedars-Sinai Medical Center under a separate project (NLST-651). The funding for this project is available under the Outstanding Investigator award R35 NIH/NHLBI grant (1R35HL161195) PI: Piotr Slomka. All the essential infrastructure is already in place in our laboratory, including necessary hardware and computing resources. Clinical oversight from physicians is also available for CAC scoring or visual CT interpretation.
Piotr Slomka Cedars-Sinai
Damini Dey Cedars-Sinai
Daniel Berman Cedars-Sinai
-
AI for Multistructure Incidental Findings and Mortality Prediction at Chest CT in Lung Cancer Screening.
Marcinkiewicz AM, Buchwald M, Shanbhag A, Bednarski BP, Killekar A, Miller RJH, Builoff V, Lemley M, Berman DS, Dey D, Slomka PJ
Radiology. 2024 Sep; Volume 312 (Issue 3): Pages e240541 PUBMED