Deep Learning Trained Without Target-Domain Labels Improves Risk Stratification Over Expert Coronary Calcium Reading on Routine Attenuation Correction CT
- 1Cedars-Sinai Medical Center
Abstract
Coronary artery calcium (CAC) is an established measure of coronary atherosclerosis on computed tomography (CT). Many SPECT and most PET myocardial perfusion studies already include a low-dose CT scan from which coronary calcium could be reported, yet it almost never is. CAC may not be reported due to image noise, cardiac motion, or the absence of a reliable reference standard. Deep learning (DL) models trained on these low-dose CT scans inherit the same uncertainty, because their reference labels are drawn on the degraded images. We bypassed this limitation by training a DL CAC model entirely outside the target domain: expert labels from high-quality gated CT were inherited by synthetic ungated images derived from those same scans, so that no CTAC scan was annotated, and no real CTAC image was used, at any point in training. Models were trained using one center and externally tested in two other centers. Because each synthetic image is derived from the gated scan itself, expert gated segmentations remain valid voxel-for-voxel without registration. Using these pairs, we developed a cross-domain knowledge adaptation model (knowledge adaptation, KAD-CAC). The DL models were externally tested on 5969 patients with a median age of 64 (IQR 56 - 73), of whom 50.2% were male. The KAD-CAC model had higher Cohens kappa K (0.86, 95% CI 0.84 - 0.88) compared to a convolutional LSTM model (K 0.78, 95% CI 0.76 - 0.80, p<0.01), or models trained with only gated images (K 0.81, 95% CI 0.80 - 0.82, p<0.01). Net reclassification improvement for death or myocardial infarction over a clinical and perfusion baseline was greatest for the KAD-CAC model and exceeded that of expert CAC reading on the same CTAC scans (difference 0.089, 95% CI 0.051 to 0.125, p<0.001). A model trained without any labels from CTAC scans therefore stratified risk of death or MI better than expert reading of the same scans. Patients classified as CAC 0 by KAD-CAC had lower adjusted risk of death or MI than patients classified as CAC 0 by experts (HR 0.50, 95% CI 0.25 - 0.98). We use paired synthetic ungated scans to transfer expert gated CAC annotations into the ungated domain, resulting in better vessel-level CAC scoring and improved risk reclassification compared with expert reading of the same scans. Automated CAC scoring could be added to every myocardial perfusion imaging report without additional imaging or radiation. The approach provides a general route to expert-quality supervision for any acquisition in which the native reference standard is unreliable.
Related
- Perivascular fat attenuation index improves the identification of functional myocardial ischemia beyond traditional risk factors: a coronary CT angiography-derived fractional flow reserve study
- Machine learning‑based risk stratification for moderate‑to‑severe anxiety‑depression in patients with hypertension and/or coronary heart disease: a multicenter cross‑sectional study
- Test-bolus peak attenuation for predicting aortic enhancement in coronary computed tomography angiography: an exploratory subgroup analysis
- Prevalence and Clinical Correlates of Coronary Perivascular Inflammation in an Asymptomatic Adult Primary Prevention Population: An Assessment of the CT-Derived Fat Attenuation Index Score (FAI-Score)
- Artificial Intelligence in Acute Coronary Syndromes: From Early Diagnosis to Culprit Lesion Characterization, Coronary Physiology, and Risk Stratification
BibTeX
@misc{shanbhag2026deep,
title = {{Deep Learning Trained Without Target-Domain Labels Improves Risk Stratification Over Expert Coronary Calcium Reading on Routine Attenuation Correction CT}},
author = {A. Shanbhag and R. J. Miller and A. Killekar and A. M. Marcinkiewicz and J. Zhou and M. Lemley and A. Kamagate and S. D. Van Kriekinge and P. B. Kavanagh and A. Feher and E. J. Miller and J. X. Liang and D. S. Berman and D. Dey and R. M. Leahy and P. Slomka},
howpublished = {medRxiv},
year = {2026},
month = jul,
doi = {10.64898/2026.07.02.26356002},
url = {https://www.medrxiv.org/content/10.64898/2026.07.02.26356002v2},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - Deep Learning Trained Without Target-Domain Labels Improves Risk Stratification Over Expert Coronary Calcium Reading on Routine Attenuation Correction CT
AU - A. Shanbhag
AU - R. J. Miller
AU - A. Killekar
AU - A. M. Marcinkiewicz
AU - J. Zhou
AU - M. Lemley
AU - A. Kamagate
AU - S. D. Van Kriekinge
AU - P. B. Kavanagh
AU - A. Feher
AU - E. J. Miller
AU - J. X. Liang
AU - D. S. Berman
AU - D. Dey
AU - R. M. Leahy
AU - P. Slomka
PY - 2026
DA - 2026/07/02
PB - medRxiv
DO - 10.64898/2026.07.02.26356002
UR - https://www.medrxiv.org/content/10.64898/2026.07.02.26356002v2
AB - Coronary artery calcium (CAC) is an established measure of coronary atherosclerosis on computed tomography (CT). Many SPECT and most PET myocardial perfusion studies already include a low-dose CT scan from which coronary calcium could be reported, yet it almost never is. CAC may not be reported due to image noise, cardiac motion, or the absence of a reliable reference standard. Deep learning (DL) models trained on these low-dose CT scans inherit the same uncertainty, because their reference labels are drawn on the degraded images. We bypassed this limitation by training a DL CAC model entirely outside the target domain: expert labels from high-quality gated CT were inherited by synthetic ungated images derived from those same scans, so that no CTAC scan was annotated, and no real CTAC image was used, at any point in training. Models were trained using one center and externally tested in two other centers. Because each synthetic image is derived from the gated scan itself, expert gated segmentations remain valid voxel-for-voxel without registration. Using these pairs, we developed a cross-domain knowledge adaptation model (knowledge adaptation, KAD-CAC). The DL models were externally tested on 5969 patients with a median age of 64 (IQR 56 - 73), of whom 50.2% were male. The KAD-CAC model had higher Cohens kappa K (0.86, 95% CI 0.84 - 0.88) compared to a convolutional LSTM model (K 0.78, 95% CI 0.76 - 0.80, p<0.01), or models trained with only gated images (K 0.81, 95% CI 0.80 - 0.82, p<0.01). Net reclassification improvement for death or myocardial infarction over a clinical and perfusion baseline was greatest for the KAD-CAC model and exceeded that of expert CAC reading on the same CTAC scans (difference 0.089, 95% CI 0.051 to 0.125, p<0.001). A model trained without any labels from CTAC scans therefore stratified risk of death or MI better than expert reading of the same scans. Patients classified as CAC 0 by KAD-CAC had lower adjusted risk of death or MI than patients classified as CAC 0 by experts (HR 0.50, 95% CI 0.25 - 0.98). We use paired synthetic ungated scans to transfer expert gated CAC annotations into the ungated domain, resulting in better vessel-level CAC scoring and improved risk reclassification compared with expert reading of the same scans. Automated CAC scoring could be added to every myocardial perfusion imaging report without additional imaging or radiation. The approach provides a general route to expert-quality supervision for any acquisition in which the native reference standard is unreliable.
ER -