Evaluation of AI-based GenStrain in Acute Myocardial Infarction

Presenting Author(s)

• 

  Sona Ghadimi, PhD

  research scientist
  University of Virginia
  Charlottesville, Virginia, United States

Primary Author(s)

• 

  Sona Ghadimi, PhD

  research scientist
  University of Virginia
  Charlottesville, Virginia, United States

Co-Author(s)

• RP

  Rishov Paul

  Graduate Research Assistant
  University of Virginia, United States
• DB

  Derek Bivona, PhD

  Data scientist
  University of Virginia Health System, United States
• RH

  Rohan Herur

  MD Candidate
  University of Virginia, United States
• CD

  Cristiane C. De Carvalho Singulane, MD

  Postdoctoral Researcher
  University of Virginia Health System
  Charlottesville, Virginia, United States
• SW

  Shuo Wang, MD, PhD

  Research Associate
  University of Virginia Health System
  Charlottesville, Virginia, United States
• KM

  Kenneth Mangion

  Research Fellow, Honorary Clinical Senior Lecturer
  University of Glasgow, United Kingdom
• 

  Colin Berry, MD, PhD

  Professor of Cardiology
  University of Glasgow
  Glasgow, Scotland, United Kingdom
• 

  Kenneth Bilchick, MD, MSc

  Associate Professor of Medicine
  University of Virginia Health System
  Charlottesville, Virginia, United States
• 

  Amit R. Patel, MD, FSCMR

  Professor of Medicine
  Division of Cardiology, University of Virginia Health System, Charlottesville, Virginia, USA.
  Charlottesville, Virginia, United States
• MZ

  Miaomiao Zhang, PhD

  Assistant Professor
  University of Virginia
  Charlottesville, Virginia, United States
• 

  Frederick H. Epstein, PhD

  Professor
  University of Virginia
  Charlottesville, Virginia, United States

Background: While strain analysis of CMR cine images is typically performed using feature tracking (FT), artificial intelligence methods based on supervised learning have recently been developed and show great potential^1,2. For example, StrainNET, a model trained on DENSE data, showed better strain agreement than FT with DENSE when applied to a diverse testing dataset¹. We evaluated a redesigned StrainNET method based on the diffusion model (GenStrain)³ in the context of myocardial infarction (MI) by examining the relationships between circumferential strain as assessed by both GenStrain and FT and metrics of late gadolinium enhancement (LGE).

Methods: CMR data from 46 patients with acute MI were used. The study had ethics approval and all subjects provided informed consent. CMR was performed at 1.5T and included short-axis cine bSSFP imaging covering the entire left ventricle (LV), DENSE imaging at basal, mid-level, and apical slices, and short-axis late gadolinium enhancement (LGE) imaging covering the entire LV. Circumferential strain analysis of short-axis cine images was performed using both GenStrain and FT (Neosoft), where both methods utilized common LV contours developed using the FT software. Strain analysis of DENSE was performed using the Open Source DENSEanalysis code^4,5. LGE images were also analyzed using Neosoft software.

Results: The dataset yielded 117 slices with high-quality cine, DENSE, and LGE images, and 327 slices with high-quality cine and LGE images. Using the high-quality cine and LGE images, Fig. 1 shows that GenStrain distinguished differences in strain between segments with 0%, 1-50%, and >50% transmural extent of infarction; whereas FT only distinguished differences in strain between segments with >50% transmural extent of infarction vs. all others. Fig. 1 also shows less strain variance using GenStrain compared to FT for each transmural extent of LGE. Across all LGE-defined groups (0%, ≤50%, >50%), StrainNet demonstrated significantly lower variance than FT for both whole-slice and segmental end-systolic strain measurements (p < 0.05 for all comparisons, Pitman–Morgan test), indicating greater consistency in strain estimation. Using the high-quality DENSE, cine, and LGE images, when correlating whole-slice strain vs. the percentage of the slice demonstrating LGE, the R² values were 0.34, 0.33, and 0.26 for DENSE, GenStrain, and FT, respectively (Fig. 2). Fig. 3 shows an example LGE image and end-systolic strain maps at a matched location obtained by DENSE, GenStrain, and FT from one patient.

Conclusion: For segmental strain analysis, GenStrain showed less strain variability and a greater ability to distinguish between grades of LGE transmural extent than FT. For whole-slice strain, GenStrain showed a correlation similar to that of DENSE and greater to that of FT with the percentage of LGE per slice. GenStrain, a diffusion model trained using DENSE data, has the potential to outperform FT for the detection of injured myocardium.

Fig. 1. Segmental end-systoic circumferential strain assessed by GenStrain and feature tracking (FT) for various LGE extents in myocardial infarction patients. Variance is lower for GenStrain vs FT, resulting in the ability of GenStrain to distinguish strain differences between 0% and 1-50%.
Fig. 2. Linear correlations between end-systolic whole-slice circumferential strain and the percent of late gadolinium per slice for (A) DENSE, (B) GenStrain and (C) FT, with R2 values of 0.34, 0.33, and 0.26, respectively.
Fig. 3. Example LGE image (A) and map (B), and end-systolic strain maps at a matched location obtained by FT (C), GenStrain (D), and DENSE (E) from one patient.