Free-breathing single-beat exercise cardiovascular magnetic resonance with generative artificial intelligence for evaluation of volumetric and functional cardiac indices: A reproducibility study

Figures

• 

  Fig. 1. (A) Free-breathing, ECG-triggered, single-beat cine acquisition collects data for each slice in one heartbeat. A dummy heartbeat is used to take the magnetization to steady-state. (B) Image reconstruction pipeline with non-uniform k-space sampling. Low-resolution images are enhanced with REGAIN. REGAIN processes each image in 32 ms, with data transfer between the external server and scanner console taking 4.6 ms per image. (C) Exercise protocol using identical cine imaging for rest and stress imaging. ECG electrocardiogram-gated, REGAIN resolution-enhancement generative adversarial inline neural network, CMR cardiovascular magnetic resonance, CS compressed sensing, iFFT inverse Fast Fourier Transform.

• 

  Fig. 2. (A) Reconstruction pipeline: data from Ex-CMR is sent to the Framework for Image Reconstruction (FIRE) and then to an external server for AI processing. Reconstructed images are returned to the scanner console in DICOM format for immediate display. (B) Image analysis workflow: preprocessing and quality control ensure consistency in cardiac phases and enable precise rest-stress comparisons. LV and RV basal/apical slices are identified for full biventricular coverage, particularly in stress images where post-exercise displacement may occur. End-systole and end-diastole are manually designated, and selected images are transferred to a new series for alignment and analysis, with two frames per slice. Rigorous quality control confirms the feasibility of endocardial tracing, complete cycle capture, and absence of PVCs, ensuring matched slices and minimal respiratory variation. EDV end-diastolic volume, EF ejection fraction, ESV end-systolic volume, LV left ventricle, PVC premature ventricular contraction, RV right ventricle, SV stroke volume, Ex-CMR exercise cardiovascular magnetic resonance, DICOM digital imaging and communications in medicine

• 

  Fig. 3. Four representative examples of image quality assessment in free-breathing single-beat acquisition are presented. Endocardial borders (white filled arrows) and epicardial borders (narrow white arrows) are emphasized. In images labeled as good (panels A and B), well-defined borders are observed at both endocardial and epicardial levels. In images in panel B, despite the chemical shift artifact close to the epicardial border, delineation of the epicardial border is feasible. In images labeled as adequate (panel C), blurring intensity increases, mainly disturbing epicardial tracing. However, endocardial tracing, despite blurring, remains feasible, allowing accurate volumetric assessment. In images labeled as poor (panel D), both epicardial and endocardial tracing are impaired

• 

  Fig. 4. Example of rest (HR = 59 bpm) and stress imaging (HR = 88 bpm) using single-beat cine acquisition in a patient with heart failure, illustrating three representative slices acquired from the basal, mid-ventricular, and apical regions. Minimal blurring in stress single-beat cine is attributed to cardiorespiratory motion artifacts. HR heart rate

• 

  Fig. 5. Qualitative assessment. (A) Diagnostic and (B) image quality assessments of both rest and stress cine acquisitions across all subjects (n = 136). Four independent readers rated the data using a 3-point Likert scale for (A) diagnostic quality (1: diagnostic; 2: non-confident; 3: non-diagnostic) and a 3-point Likert scale for (B) image quality (1: good; 2: adequate; 3: poor). Note: For values less than 5%, only color-coding is provided

• 

  Fig. 6. Inter-observer agreement for LV and RVEF obtained from rest and stress imaging in 41 (30%) of the acquired data. The top row shows regression plots illustrating the agreement between both readers. The bottom row displays Bland-Altman plots of the bias between the two readers. EF ejection fraction, LV left ventricular, MD mean difference, RV right ventricular, SD, standard deviation

• 

  Fig. 7. Regression plots of LV volumetric measurements agreement between visits 1 and 2 in two readers. Displayed are regression plots of the inter-visit agreement LVEDV, LVESV, and LVEF measurements at rest (top row) and during exercise stress (bottom row). Red indicates measurements by reader 1, blue indicates measurements by reader 2. EDV end-diastolic volume, EF ejection fraction, ESV end-systolic volume, LV left ventricular

• 

  Fig. 8. Regression plots of agreement of RV volumetric measurements between visits 1 and 2 in two readers. Displayed are regression plots of the inter-visit agreement RVEDV, RVESV, and RVEF measurements at rest (top row) and during exercise stress (bottom row). Red indicates measurements by reader 1, blue indicates measurements by reader 2. EDV end-diastolic volume, EF ejection fraction, ESV end-systolic volume, RV right ventricular

• 

  Fig. 9. Measurements of biventricular (LV/RV) stroke volume (SV) and ejection fraction (EF) for Rest1, Rest2, and Stress imaging during one visit, illustrating measurable volumetric and functional changes triggered by exercise, with an MPHR of 67 ± 10%. MPHR maximum age-predicted heart rate, LV left ventricular, RV right ventricular