Synthetic Contrast-Free LGE via Diffusion-Based Framework in Acute MI for Image Quality and Quantitative Scar Analysis.
Authors
Affiliations (8)
Affiliations (8)
- Medical Big Data Research Center, Medical Innovation Research Division of PLA General Hospital, Beijing, China (J.Q., X.Y., M.H., J.L., K.H.).
- Chinese PLA Medical School, The First Medical Center, Chinese PLA General Hospital, Beijing. (J.Q., X.Y., M.H., J.L.).
- Department of Radiology, The First Medical Center, Chinese PLA General Hospital, Beijing. (J.C., Y.Z., T.L.).
- Department of Radiology, Beijing Jishuitan Hospital, Capital Medical University, China (J.C.).
- Department of Radiology, Central Hospital Affiliated to Shandong First Medical University, Jinan, China (J.W.).
- Department of Radiology, Zhongshan Hospital, Fudan University, Shanghai, China. (Y.C., H.J.).
- Department of Medical Imaging, Shanghai Medical School, Fudan University and Shanghai Institute of Medical Imaging, China (Y.C., H.J.).
- Human Phenome Institute and Shanghai Pudong Hospital, Fudan University, Shanghai, China. (C.W.).
Abstract
This study aims to develop a diffusion model-based framework for generating late gadolinium enhancement (LGE)-like images without contrast. The resulting synthetic images are then comprehensively evaluated for subjective and objective image quality, as well as their clinical utility for quantifying scar in acute myocardial infarction. In this retrospective study, we developed a diffusion mode-based framework, multisequence guided diffusion to generate synthetic native enhancement (SNE) images from cine magnetic resonance imaging, and T2 short tau inversion recovery sequences. Data were collected from 331 patients with acute myocardial infarction across 3 centers from January 2014 to July 2024. Subjective and objective image qualities were assessed using Likert scoring and contrast ratio analyses on both internal and external cohorts, comparing SNE with standard LGE to evaluate group differences. Myocardial contours were manually delineated, and scar size and transmurality were quantified using the full-width at half-maximum method to assess the accuracy of myocardial infarction detection. In comparisons with general generative models and multimodal fusion-based generative approaches, multisequence guided diffusion demonstrated more favorable visual perceptual quality and the closest data distribution alignment to conventional LGE. SNE demonstrated significantly higher quality than LGE (internal: 4.250 [4.000-4.750] versus 4.000 [3.750-4.500]; external: 4.250 [4.000-4.750] versus 4.000 [3.500-4.250]; <i>P</i><0.05) and improved contrast ratios (blood pool versus myocardium: 9.010 [6.938-12.761] versus 8.767 [6.361-11.745] internally and 16.871 [12.546-24.237] versus 13.472 [9.380-19.599] externally; <i>P</i><0.05). SNE showed strong agreement with LGE for scar size (internal <i>R</i>=0.839; external <i>R</i>=0.816; <i>P</i><0.001) and transmurality (internal <i>R</i>=0.792; external <i>R</i>=0.758; <i>P</i><0.001) with minimal biases (scar size: 2.490% internal, 2.222% external; transmurality: 2.984% internal, 2.225% external), indicating accurate scar depiction and robust generalizability. SNE demonstrated strong agreement with LGE in quantitative assessment of acute myocardial infarction scar, with comparable or improved image quality.