Deep Learning-based Post Hoc CT Denoising for the Coronary Perivascular Fat Attenuation Index

Area of Science:

• Cardiovascular Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Coronary inflammation in high-risk plaques can be assessed using the fat attenuation index (FAI) via coronary computed tomography angiography (CCTA).
• FAI is sensitive to image noise, potentially limiting its diagnostic accuracy.
• Deep learning (DL) offers a promising approach for post hoc noise reduction in CCTA.

Purpose of the Study:

• To evaluate the diagnostic performance of FAI in DL-based denoised CCTA images.
• To compare the diagnostic capability of denoised FAI with that of coronary plaque magnetic resonance imaging (MRI) for identifying high-intensity hemorrhagic plaques (HIPs).

Main Methods:

• Retrospective analysis of 43 patients who underwent both CCTA and coronary plaque MRI.
• Generation of high-fidelity CCTA images using a DL residual dense network for noise reduction.
• Measurement of FAI on original and denoised CCTA images.
• HIPs on MRI served as the diagnostic reference standard.
• Assessment of diagnostic performance using receiver operating characteristic curves.

Main Results:

• DL-based denoising significantly improved the area under the curve (AUC) for FAI in predicting HIPs (0.89 vs. 0.77, p=0.008).
• The optimal cutoff for FAI in denoised CCTA achieved 85% sensitivity and 79% specificity for HIP detection.
• Overall accuracy for HIP prediction using denoised FAI was 80%.

Conclusions:

• DL-based noise reduction enhances the diagnostic performance of FAI in CCTA.
• Denoised high-fidelity CCTA improves the accuracy and specificity of FAI for predicting high-risk hemorrhagic plaques.