A deep learning model may offer the promise of fully automated quantification of coronary plaque based on assessment of coronary computed tomography angiography (CCTA) scans.
For the retrospective study, recently published in Radiology, researchers developed and validated the fully automated deep learning model PlaqueSegNet for quantifying plaque volume. The training set was comprised of 1,409 patients (mean age of 63) and an internal validation set consisted of 604 patients (mean age of 63), according to the study. The study authors noted that 61 percent of the total cohort had stable angina and 30 percent had unstable angina.
Overall, across four external datasets, the deep learning model achieved greater than 90 percent intraclass correlation coefficient (ICC) agreement on plaque volume (PV) assessment with expert readers and intravascular ultrasound (IVUS) evaluation.
Specifically, among paired CCTA and intravascular ultrasound (IVUS) samples, the researchers noted that PlaqueSegNet offered a 93 percent ICC agreement with IVUS for PV assessment, and no significant differences with respect to calcified, non-calcified or mixed PV. In another subgroup, PlaqueSegNet provided a 94 percent ICC with expert readers on PV assessments, and the study authors noted no discernible differences across computed tomography (CT) scanners from multiple manufacturers.
In a cohort of 10,644 patients with non-obstructive coronary artery disease (CAD), the researchers found that those with the highest quartile of total PV assessed with PlaqueSegNet had an adjusted hazard ratio of 2.25 for the development of major adverse cardiac events (MACE) in comparison to patients with the lowest quartile of total PV.
“These results reinforce the role of fully automated PV quantification in enhancing risk stratification, particularly in the clinically prevalent non-obstructive CAD population,” noted lead study author Qian Chen, MD, who is affiliated with the Department of Radiology at Nanjing First Hospital at Nanjing Medical University in Nanjing, China, and colleagues.
Three Key Takeaways
• High agreement with expert and IVUS standards. The fully automated deep learning model (PlaqueSegNet) demonstrated >90 percent intraclass correlation coefficient (ICC) for plaque volume (PV) quantification across external datasets, including 93 percent ICC agreement with intravascular ultrasound (IVUS) and 94 percent with expert readers, with consistent performance across plaque subtypes and CT vendors.
• Strong prognostic value in non-obstructive CAD. In over 10,000 patients with non-obstructive CAD, higher total plaque volume was independently associated with increased risk of major adverse cardiac events (MACE) (HR 2.25 for highest vs. lowest quartile), supporting automated PV as a robust risk stratification tool.
• Utility for longitudinal monitoring and risk prediction. Serial CCTA analysis showed plaque progression detected by the model predicted MACE (C-index 74 percent) with a 2.52-fold higher event risk, highlighting the potential for guiding individualized therapy and tracking disease progression over time.
The study authors also determined that in a cohort of 270 patients who had serial CCTA exams with a median follow-up of 3.6 years, plaque progression detected with PlaqueSegNet had a 74 percent C-index for predicting MACE. This analysis also revealed that those with plaque progression were 2.52 times more likely to develop MACE, according to the study authors.
“To our knowledge, our study is the first to assess the reproducibility of a DL-based model for plaque quantification across serial CCTA examinations and demonstrate the potential of a DL-based model for monitoring PV changes and guiding individualized therapy,” added Chen and colleagues.
(Editor’s note: For related content, see “What New CCTA Research Reveals About Quantitative Plaque Assessment and Predicting MACE,” “Where Do Things Stand with AI-Powered Plaque Quantification for CCTA Exams?: An Interview with Ron Blankstein, MD” and “Why Plaque Burden is Critical to Assessing Cardiovascular Risk: An Interview with Ibrahim Danad, MD, PhD.”)
In regard to study limitations, the authors acknowledged the lack of coronary stenosis assessment for PlaqueSegNet, the exclusion of poor quality CCTA images and the use of a small retrospective cohort for prognostic validation of the deep learning model in patients who had serial CCTA exams.
