Deep learning model for low-dose CT late iodine enhancement imaging and extracellular volume quantification

Area of Science:

• Cardiovascular Imaging
• Artificial Intelligence in Medical Imaging
• Radiology

Background:

• Late iodine enhancement (LIE) imaging is crucial for assessing myocardial scar and extracellular volume (ECV) in cardiovascular CT.
• Traditional LIE images can suffer from noise, potentially affecting diagnostic accuracy and quantification.
• Deep learning (DL) offers a promising approach to enhance image quality and improve quantitative analysis.

Purpose of the Study:

• To develop and validate DL models for denoising LIE images.
• To assess the impact of denoised LIE images on the accuracy of ECV quantification.
• To compare the performance of different DL models (RDN and cGAN) in image denoising and quantitative analysis.

Main Methods:

• Retrospective analysis of LIE images from 423 patients undergoing cardiac CT.
• Development and validation of two DL models: residual dense network (RDN) and conditional generative adversarial network (cGAN).
• Comparison of image quality metrics (SNR, CNR) and diagnostic performance (AUC) for LIE and ECV quantification between original and denoised images (LIEsingle, LIEaveraging, LIE_RDN, LIE_GAN).

Main Results:

• Both RDN and cGAN models significantly improved image quality (SNR, CNR) compared to original LIE images.
• The RDN model demonstrated superior performance, yielding higher SNR and CNR values.
• LIE_RDN images significantly improved the area under the curve (AUC) for LIE evaluation and enabled accurate ECV quantification compared to other methods.

Conclusions:

• The RDN deep learning model effectively denoises LIE images, markedly enhancing SNR and CNR.
• Denoised LIE images using the RDN model improve visual analysis identifiability.
• Accurate CT-derived ECV quantification can be achieved using single-stack LIE images denoised by the RDN model.