Skip to main content
SCMR 29th Annual Scientific Sessions Main banner
Final Program


Program-At-A-Glance

Oral Abstracts Session

Virtual Recording

Session: Oral Abstracts Clinical Session 1

CMR in mixed-use MRI sites with no CMR training using AutoCMR, an automated 3D exam

Thursday, February 5, 2026
8:00 AM - 8:10 AM  
Location: Asia IV and Hall 8 (3rd Floor)
Background: Despite the demonstrated importance of cardiac magnetic resonance (CMR) in cardiovascular care1–3, CMR access remains limited4,5. While non-cardiac MRI is widely performed at community imaging centers, CMR is largely restricted to academic hospitals4,5. The expansion of cardiac MRI into mixed-use environments is hindered by the need for specially trained CMR technologists and long scan times compared to non-cardiac indications6,7. With the goal of improving CMR access, we proposed AutoCMR, a simple 30-minute exam providing 3D isotropic cine and delayed enhancement (DE) imaging and automated analysis8. In this work, we demonstrate the feasibility of AutoCMR at two community health centers (CHC) with mixed-use MRI programs and no CMR technologists.

Methods: 46 patients with clinical CMR in the past year were scanned at one of two CHCs with IRB approval (Fig 1a). At both sites AutoCMR was performed on a 1.5T Altea (Siemens Healthineers), including a < 1 minute pre-scan, 10-12 minutes of cine (TR/TE=4.4/2.9ms, FA=15°), and 15 minutes of DE (pulse-ox gated, TR/TE=4.4/2.8ms, FA=10°, 900 heartbeats) using Gaussian random Cartesian sampling (GRE, 307mm FOV, 1.6x1.6x1.6mm3, Gmax=22 mT/m, slewmax=100mT/m/ms). Cine images were automatically segmented using a Swin-UNETR model11 and used to calculate 4-chamber function metrics (Fig 1e). End diastolic volume (EDV) and ejection fraction (EF) in the left ventricle (LV), right ventricle (RV), left atrium (LA), and right atrium (RA) from AutoCMR (ACMR) were compared to previously acquired clinical values (2DCMR) using Bland-Altman analysis and intraclass correlation. Image quality and diagnostics were evaluated in a subset of 8 patients by two Level 3 CMR readers.

Results: Initial sequence development was performed in 10 patients. An additional 6 subjects were excluded due to invalid positioning (2), automated self-gating (1), and automated segmentation (3). In the final cohort of 30 volunteers, total scan time was 26.8±2.9 minutes (Fig 1d). AutoCMR images acquired at both CHC sites by technologists with no CMR training are qualitatively comparable to clinical CMR (Fig 2a-b). AutoCMR image quality was rated slightly lower than clinical 2DCMR (cine: 3.55 vs 3.76, DE: 3.75 vs 3.9), with more noise but fewer susceptibility artifacts (Fig 2c). AutoCMR LV volume and EF are correlated (ICC=0.92, 0.69) with clinically reported values and have low bias (bias=-8.49mL, 2.82%) (Fig 3a-b). RV, LA, and RA volume and EF are also correlated (ICC > 0.66, 0.53). Atrial volume bias is attributed to the calculation of volume based on a single 4CH image for 2DCMR. Diagnostic evaluation of AutoCMR and 2DCMR images yielded generally reduced but comparable accuracy compared to clinical reports (Fig 3c).

Conclusion: With its push-button, 30-minute design, AutoCMR is feasible in mixed-use MRI sites. By lowering technical barriers and providing consistent scan time, AutoCMR has the potential to enable routine CMR in hospital settings beyond major academic institutions.

Figure 1. 46 Patients were scanned with AutoCMR at community health centers with no CMR technologists. a, All patients received clinical 2D CMR exams at an academic hospital within 1 year of their AutoCMR exam, performed at one of two CHCs. b-c, AutoCMR is a push-button exam providing 3D cine and DE images with spoiled gradient echo acquisitions. d, AutoCMR images are reconstructed using localized self-navigation. Cine images are automatically segmented across all 30 frames, and segmentations are used to generate volume and function metric summaries. e, In 30 patients, total AutoCMR scan time was significantly shorter than clinical exam time adjusted to include only localizers, cine, and DE imaging (ACMR: 26.8±2.9 minutes, 2DCMR: 34.1±10.5 minutes, p=0.0006), with the longest AutoCMR exam taking 35 minutes, while 11 of 30 clinical exams took 35 minutes or longer.
Figure 2. AutoCMR images acquired at community health centers are comparable to clinical 2D CMR images acquired by CMR technologists at an academic hospital. a-b, Representative AutoCMR images (top) acquired at two different community health centers are compared to 2D CMR images (bottom). In (a), the patient with suspected cardiac sarcoidosis is observed to have nearly transmural delayed enhancement in the basal to mid septum. The repaired tetralogy of Fallot patient shown in (b) is observed to have significant metal artifacts in the clinical 2DCMR images, which are reduced in the GRE-based AutoCMR images. No delayed enhancement is observed. c, In a subset of 8 patients reviewed by 2 level 3 CMR readers, both cine and DE AutoCMR images are noisier than clinical 2D images (cine: 2.50 vs 3.69, DE: 3.00 vs 3.75), but are otherwise rated comparably (overall cine: 3.55 vs 3.76, overall DE: 3.75 vs 3.90).
Figure 3. CMR metric extraction and diagnostic review are feasible for AutoCMR performed at a community health center. a, Ventricular volumes extracted from automatic segmentations of AutoCMR cine images demonstrate good reliability in intraclass correlation (LV: 0.92, RV: 0.90; 95% confidence interval shown in gray) and low bias in Bland Altman analysis (LV: -8.49mL, RV: 14.9mL) compared to reported values from previously acquired clinical 2DCMR exams. Atrial AutoCMR volumes are moderately correlated with clinical 2D volumes (ICC ≥ 0.66) with bias reflective of the uncertainty in calculating atrial volumes from a single clinical 2D image. b, AutoCMR ejection fraction is moderately correlated in all four chambers (ICC ≥ 0.53), with bias <4% in the LV, RV, and RA. c, In a diagnostic review of 2DCMR and AutoCMR images compared to clinical MRI reports, accuracy in identifying late gadolinium enhancement was comparable for AutoCMR and clinical 2D images, with (0.63, 0.69) accuracy in identifying LV LGE positive cases and (0.56, 0.63) accuracy in identifying LGE negative cases, respectively. LV: left ventricle, RV: right ventricle, LA: left atrium, RA: right atrium.