PQA 03 - Digital Health Innovation and Informatics, Patient Safety & Quality, and Radiation and Cancer Biology
Presenter(s)
C. Ciausu1, M. P. C. Tonneau2, A. Warrington3, C. V. Guthier3, K. M. Atkins4, H. Aerts3, and R. H. Mak5; 1Artificial Intelligence in Medicine (AIM) Program, Mass General Brigham, Harvard Medical School, Boston, MA, 2Mass General Brigham AIM, Boston, MA, 3Department of Radiation Oncology, Mass General Brigham/Dana-Farber Cancer Institute, Harvard Medical School, Boston, MA, 4Department of Radiation Oncology, Cedars-Sinai Medical Center, Los Angeles, CA, 5Department of Radiation Oncology, Brigham and Women’s Hospital, Dana-Farber Cancer Institute, Harvard Medical School, Boston, MA
Purpose/Objective(s):
AI-based coronary artery calcium (CAC) detection is increasingly used for opportunistic cardiovascular risk assessment on non-contrast, diagnostic chest CT. In thoracic radiation oncology workflows, however, a usable non-contrast free-breathing (FB-CT) chest CT may be unavailable because the available chest CT is contrast-enhanced, whereas non-contrast four-dimensional (4D) CT are routinely acquired for simulation. Respiratory motion can alter the quantitative visualization of calcium, so phase choice may affect AI-derived CAC. We evaluated which 4D phase correlates best with FB-CT and diagnostic CT CAC and hypothesize that the 40% phase (end expiration) is a practical surrogate for a non-contrast chest CT.
Materials/Methods:
The internal phase-selection cohort from 2021-2023 included 808 lung cancer patients treated with radiotherapy, and retaining only patients (n=332) with simulation FB-CT and all 10 phases of 4D-CT scans. AI-based contrast detection was applied to FB-CT scans to eliminate contrast cases, and a previously validated AI-based CAC algorithm was run on all images in the resulting non-contrast cohort (n=265). AI-CAC scores were median-aggregated per patient/phase and compared with FB-CT scans and evaluated using ordinal Agatston risk groups. The 4D-CT phase versus FB-CT scan comparisons were further analyzed using Spearman correlation and median absolute error (MAE) for AI-predicted Agatston score, and quadratic weighted kappa (QWK) for Agatston risk group. For validation, we used a 2023-2024 cohort (n=90) with a breath-hold, diagnostic CT scan from pre-treatment PET-CT as reference. Results:
CAC scores from FB-CT were strongly skewed with a high-score tail (median Agatston score: 383.5, IQR 1415.3), and 2.5%, 14.4%, 25.0%, and 48.1% were in the 0, 1-100, 101-400 and >400 risk groups respectively. the phase-selection cohort, the highest correlation with FB-CT was at the 10% phase (rho=0.733), while 40% phase was a close second (rho=0.727) and had the best median absolute error (225 Agatston units). Using a 5-level Agatston risk group, internal cohort 40% versus FB-CT concordance was: exact agreement 45.2% (95% CI 35.6–54.8%), adjacent-or-exact agreement 72.1% (95% CI 63.5–80.8%), and QWK 0.543 (95% CI 0.409–0.667). In the validation cohort, 40% versus diagnostic CT concordance was: exact agreement 53.3% (95% CI 43.3–63.3%), adjacent-or-exact agreement 75.6% (95% CI 66.7–84.4%), and QWK 0.631 (95% CI 0.496–0.749).
Conclusion:
In thoracic radiation oncology workflows where a non-contrast breath-hold CT is unavailable, the 40% 4D-CT phase provides a practical and accurate substitute for AI-based CAC cardiac screening, enabling opportunistic cardiovascular risk stratification directly from routine simulation imaging without additional scans.
