Image Quality Assessment of Deep Learning-Based Virtual Monoenergetic Images From Single-Energy CT Pulmonary Angiography.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiology, University of Wisconsin-Madison.
- Department of Medical Physics, University of Wisconsin-Madison, Madison, WI.
- Department of Imaging Physics, University of Texas MD Anderson Cancer Center, Houston, TX.
- Department of Radiology, Henry Ford Health, Detroit, MI.
Abstract
Low keV virtual monoenergetic (VME) images are effective in enhancing vessel opacification but require dual-energy CT (DECT), limiting widespread clinical use. Recent advancements in deep learning (DL) enable the generation of VME images from single-energy CT (SECT). However, the performance of the methods has not been evaluated in any clinical use case. The purpose of this work was to assess both objective and subjective image quality of deep learning-based VME images derived from heterogeneous SECT data for pulmonary angiography. In this retrospective study, 52 sets of SECT pulmonary angiography images were processed using a deep learning method to estimate material basis images. 40 keV VME images were generated from heterogeneous SECT data using a pretrained physics-constrained Deep-En-Chroma DL model. Two thoracic radiologists, blinded to the image reconstruction method, evaluated pulmonary vessel opacification and overall image quality on DL-VME and SECT images using 5-point Likert scales. Objective image quality was assessed by measuring enhanced vessel contrast and contrast-to-noise ratio (CNR). Statistical analysis was performed using paired t tests and Mann-Whitney U tests. Compared with SECT, DL-VME images demonstrated significantly higher subjective image quality score and vessel opacification score (P≤0.008). DL-VME yielded a higher average contrast for emboli (1085 vs. 331 HU, P<0.001) and improved CNR (17.8 vs. 11.1, P<0.001). Results of subgroup analysis indicate no significant variation in VME performance across patient sex, scanner model, radiation dose, and tube potential. The vessel opacification scores of both VME and SECT demonstrate dependence on patient weight, with VME providing better vessel opacity for both lighter and heavier patients. A measure of 40 keV DL-VME derived from SECT effectively enhances both vessel opacification and image quality in CT pulmonary angiography. The image quality advantage of DL-VME over SECT remains robust across variations in data acquisition and patient variables.