Learned high-resolution cardiac CT imaging from ultra-high-resolution PCD-CT

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Disease

Background:

• Coronary computed tomography angiography (cCTA) is crucial for diagnosing coronary artery disease (CAD).
• Current EID-based CT scanners have limited spatial resolution, hindering CAD evaluation due to small vessel size and calcification artifacts.
• Photon-counting-detector (PCD) CT offers ultra-high resolution (UHR) but is expensive and inaccessible.

Purpose of the Study:

• To develop and validate ILUMENATE, a super-resolution convolutional neural network (CNN).
• ILUMENATE aims to simulate UHR PCD-CT image quality from standard low-resolution (LR) cCTA data.
• To improve the visualization and diagnostic potential of cCTA for CAD patients.

Main Methods:

• A modified U-Net architecture CNN (ILUMENATE) was trained using LR and HR cCTA image patches.
• Training utilized UHR PCD-CT images reconstructed with smooth (LR input) and sharp (HR label) kernels.
• Network performance was evaluated quantitatively and qualitatively on 5 unseen patient datasets.

Main Results:

• ILUMENATE produced sharper edges, closely resembling HR reference images.
• CT number stability was maintained with <4% difference; noise was reduced by 28%.
• Structural similarity (SSIM) averaged 0.70, and percent diameter stenosis was reduced compared to input images.

Conclusions:

• ILUMENATE effectively enhances LR cCTA images, simulating UHR PCD-CT quality.
• The AI tool shows significant potential for improving CAD patient management by enhancing image quality.
• ILUMENATE offers a promising, cost-effective approach to achieving higher resolution imaging in clinical practice.