DeepCAC: Deep convolutional neural networks to predict cardiovascular risk from computed tomography
Coronary artery calcium is an accurate predictor of cardiovascular events. While it is visible on all computed tomography (CT) scans of the chest, this information is not routinely quantified as it requires expertise, time, and specialized equipment. Recent strides in artificial intelligence, deep learning in particular, have shown its viability in medical imaging applications and this makes deep learning a promising technology for automating cardiovascular event prediction from imaging.
Here, we present a deep learning system that automatically and accurately can predict cardiovascular events by quantifying the presence and extent of coronary calcium. The system was tested in an independent cohort of 20,084 individuals from four well-established prospective cohorts and randomized controlled trials - a healthy asymptomatic community-dwelling sample from the Framingham Heart Study (FHS), older asymptomatic heavy smokers in the National Lung Screening Trial (NLST), a symptomatic stable chest pain population evaluated for suspected coronary artery disease in the outpatient setting in the Prospective Multicenter Imaging Study for Evaluation of Chest Pain (PROMISE), and a symptomatic acute chest pain population presenting to the emergency department in the Rule Out Myocardial Infarction using Computer Assisted Tomography (ROMICAT-II) trial.
Overall, the association between the algorithm’s prediction and adverse cardiovascular events was tested in individuals who were imaged using different CT scanners, applying a variety of CT scan protocols, including ECG-gated and non-gated CT scans. Accuracy compared to the gold standard of expert human readers was assessed in 5,521 subjects across all four cohorts. Our results demonstrate that deep learning methods can automated cardiovascular risk predictions from medical images acquired in several clinical scenarios. These observations provide a rationale to implement this technique in both screening and hospital settings to improve population health, at high speed and low costs.
Publication
Zeleznik, R et al. Deep convolutional neural networks to predict cardiovascular risk from computed tomography. Nature Comm 12, 715 (2021). https://doi.org/10.1038/s41467-021-20966-2
Code availability
All the code of the deep learning system including the trained model are publicly available under the open MIT license and can be found here. The system consists of four consecutive steps for (1) heart localization, (2) heart segmentation, (3) coronary calcium segmentation, and (4) calcium score calculation.
Statistical Code
This link contains the code to reproduce the statistical analysis of our paper Deep convolutional neural networks to predict cardiovascular risk from computed tomography. More information about the statistical analysis can be found in the Methods section. We provide the automatically predicted as well as manual calcium scores for 396 cases from the National Lung Screening Trial (NLST). This sub-group was randomly selected from the full NLST cohort.
Data availability
Example data of computed tomography (CT) images for testing the deep learning system can be found here. For this we included four lung cancer screening thoracic CT scans where expert readers manually segmented the heart and coronary calcium and calculated a coronary calcium risk score. The image data and manual segmentations are available in the nrrd format.
Acknowledgements
We would like to thank the Framingham Heart Study, NCI, ACRIN, NLST, Prospective Multicenter Imaging Study for Evaluation of Chest Pain, and Rule Out Myocardial Infarction Using Computer Assisted Tomography II trial for access to trial data.The authors acknowledge financial support from NIH (NIH-USA U24CA194354, NIH-USA U01CA190234, NIH-USA U01CA209414, NIH-USA R35CA22052; 5R01-HL109711, NIH/NHLBI 5K24HL113128, NIH/NHLBI 5T32HL076136, NIH/NHLBI 5U01HL123339), the European Union - European Research Council (866504), as well as the German Research Foundation (DFG 1438/1-1 and 6405/2-1), American Heart Association Institute for Precision Cardiovascular Medicine (8UNPG34030172), Fulbright Visiting Researcher Grant (E0583118), Rosztoczy Foundation Grant. The Framingham Heart Study (FHS) acknowledges the support of contracts NO1-HC-25195, HHSN268201500001I and 75N92019D00031 from the National Heart, Lung and Blood Institute.