Cardiovascular

Deep Learning-based Monoenergetic Imaging for Calcified Coronary Stenosis Assessment at Energy-integrating Detector CT.

TL;DR

DIAMOND demonstrated feasibility of generating high-kiloelectron volt virtual monoenergetic images from single-energy energy-integrating detector CT, providing artifact-reduced coronary imaging and improved stenosis quantification for heavily calcified plaques comparable to photon-counting detector CT without hardware upgrades.

Key Findings

DIAMOND significantly reduced average percent diameter stenosis compared to standard EID CT, approaching values obtained with UHR PCD CT.

  • In 23 participants (mean age 69 years ± 8 [SD]; 18 male), average percent diameter stenosis (PDS) decreased from 35.65% with EID CT to 25.19% with DIAMOND (P < .05)
  • DIAMOND values approached 24.27% with UHR PCD CT (P < .05 compared to EID CT)
  • The difference between DIAMOND and PCD CT values was small (25.19% vs 24.27%)
  • Participants had heavily calcified plaques and underwent EID CT at 120 kV followed by same-day PCD CT

DIAMOND led to CAD-RADS reclassification in 42% of lesions relative to EID CT.

  • Reclassification occurred in 11 of 26 (42%) lesions based on Coronary Artery Disease Reporting and Data System (CAD-RADS) criteria
  • Reclassifications were mainly in the mild to moderate stenosis ranges
  • Changes in stenosis severity categorization were evaluated based on PDS

DIAMOND reduced blooming artifacts and improved lumen visualization with image quality resembling PCD CT.

  • DIAMOND was trained using a simplified U-Net architecture
  • 70-keV PCD VMIs served as inputs and 100-keV PCD VMIs as targets during training
  • The trained model was applied to prospective EID CT data acquired at 120 kV
  • Image quality was described as resembling PCD CT

DIAMOND processing time was approximately 0.21 second per axial section on a standard graphics processing unit.

  • Processing time of approximately 0.21 second per axial section was achieved on a standard GPU
  • This indicates practical clinical feasibility for rapid image generation

DIAMOND was trained on a small retrospective dataset of PCD CT examinations and validated prospectively.

  • Training used a retrospective dataset of only 10 CCTA examinations performed with ultrahigh-resolution PCD CT
  • The study period was August 2022 to September 2023
  • A combination of retrospective and prospective imaging data was used
  • Percent diameter stenosis was quantified for a phantom and participants using commercial software and compared across EID CT, DIAMOND, and PCD CT using Bland-Altman analysis

Have a question about this study?

Citation

Chang S, Koons E, Gong H, Thorne J, Williamson E, Foley T, et al.. (2026). Deep Learning-based Monoenergetic Imaging for Calcified Coronary Stenosis Assessment at Energy-integrating Detector CT.. Radiology. Cardiothoracic imaging. https://doi.org/10.1148/ryct.250230