Elevated Coronary Artery Calcium Quantified by a Deep Learning Model from Radiotherapy Planning Scans Predicts Mortality in Lung Cancer


      Coronary artery calcium (CAC) is one of the strongest predictors of long-term atherosclerotic coronary vascular disease in asymptomatic individuals, however the feasibility of quantitating this measurement from radiotherapy (RT) planning computed tomography (CT) scans is unknown. Since patients with non-small cell lung cancer (NSCLC) represent a distinctly high cardiovascular risk population, we sought to quantify CAC from RT planning CTs in NSCLC patients using a deep learning model.


      Retrospective analysis of non-contrast enhanced RT planning CTs from 464 consecutive locally-advanced NSCLC patients treated with thoracic RT. The CAC algorithm was previously trained on 693 independent cardiac-gated CT scans manually segmented by expert readers using three consecutive deep learning networks for segmentation and tested on three independent cohorts of 441, 664, and 398 scans. Plaques ≥1 cubic millimeter were volumetrically measured and multiplied by a maximum plaque density factor to generate an Agatson-like CAC Score. The model was used to calculate a CAC risk group for each planning CT, defined as CAC=0 (very low risk) versus CAC>0 (elevated risk). For patient factors, continuous covariates were evaluated using a Wilcoxon rank sum test whereas categorical covariates were compared using a Fisher exact test. Univariable Cox regression analysis was performed and Kaplan-Meier estimates of all-cause mortality were calculated.


      After a median follow-up of 18 months, there were 353 deaths (2-year all-cause mortality, 52.2% [95% CI, 47.7-56.8%]). Of the 464 planning CTs, 35% (162/464) were CAC=0 and 65% (302/464) were CAC>0. Patients in the CAC>0 group were older (median age 67 vs. 60 years, P<.0001), more likely male (58% vs. 37%, P<.001), have an ever-smoking history (95% vs. 84%, P<.001), and less likely treated with surgery (32% vs. 45%, P=.003) or chemotherapy (91% vs. 97%, P=.008). There was no difference in mean heart dose delivered between CAC>0 vs. CAC=0 (11.9 Gy vs. 11.5 Gy, P=.83). CAC>0 was associated with an increased risk of all-cause mortality on univariable Cox regression (hazard ratio, 1.29 [95% CI, 1.03-1.62]; P=.027). The 2-year all-cause mortality stratified by CAC group was 56.7% (95% CI, 51.1-62.3%) in CAC>0 vs. 47.1% (95% CI, 39.6-55.1%) in CAC=0 (log-rank P=.0259).


      Coronary artery calcium was effectively measured from non-contrast RT planning CTs using an automated deep learning model. Elevated CAC, as predicted by the deep neural network, was associated with an increased risk of all-cause mortality on univariable analysis in locally-advanced NSCLC patients despite a high competing risk of lung cancer death. Deeper investigation of contributing and confounding variables in a comprehensive predictive model is warranted.


      Commenting Guidelines

      To submit a comment for a journal article, please use the space above and note the following:

      • We will review submitted comments as soon as possible, striving for within two business days.
      • This forum is intended for constructive dialogue. Comments that are commercial or promotional in nature, pertain to specific medical cases, are not relevant to the article for which they have been submitted, or are otherwise inappropriate will not be posted.
      • We require that commenters identify themselves with names and affiliations.
      • Comments must be in compliance with our Terms & Conditions.
      • Comments are not peer-reviewed.