Evaluation of inline strain CMR: initial results using fully-automated DENSE and StrainNET analysis of cine images

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)

• KC

  Kelvin Chow, PhD

  MR Collaboration Scientist
  Siemens Healthcare Ltd., Canada
  Calgary, Alberta, Canada
• CD

  Cristiane C. De Carvalho Singulane, MD

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

  Parsa Pilevar

  gratuate student
  University of Virginia, United States
• JX

  Jiarui Xing

  Postdoc
  Yale University, United States
• 

  Jonathan A. Pan, MD, MSc, MBA

  Assistant Professor
  University of Virginia Health System
  Charlottesville, Virginia, United States
• CB

  Christiana M Brenin

  Associate Professor, Medicine
  University of Virginia cancer center, United States
• 

  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: Strain analysis in cardiac magnetic resonance (CMR) is traditionally performed offline and typically requires manual contouring by an expert, is time-consuming, and does not facilitate high-throughput clinical workflows. Automated inline strain analysis offers the potential to streamline clinical integration. We previously developed automatic methods for strain analysis of displacement encoding with stimulated echoes (DENSE)¹—a reference standard for myocardial strain—and for cine CMR using StrainNET^2,3. Here we evaluated inline implementations of both methods.

Methods: Inline strain analysis for cine and DENSE was deployed on a 3T scanner (MAGNETOM Prisma, Siemens, Germany) using the Framework for Image Reconstruction Environment (FIRE)⁴ research prototype. Images were reconstructed in the Image Calculation Environment and automatically transferred to a containerized Python environment for analysis. The Python pipeline included 3D (2D + time) left ventricular (LV) segmentation via deep learning, phase unwrapping for DENSE, Lagrangian displacement estimation via StrainNET³ for cine or a regularized least squares method⁴ for DENSE, circumferential strain calculation, strain map visualization, RV–LV insertion point extraction, and calculation of whole-slice and segmental strain-time curves with end-systolic strain quantification.

24 patients (13 female undergoing cardiotoxicity screening and 11 (1 female) with heart failure) underwent CMR with bSSFP cine and DENSE. Inline strain analysis was automatically performed immediately post-acquisition, and offline user-interactive DENSE analysis⁶ served as the reference standard. Agreement between inline and offline methods was assessed using Bland–Altman and Pearson correlation analyses for end-systolic circumferential strain.

Results: Inline analysis produced strain maps and strain-time curves within 15 seconds of image acquisition for both DENSE and cine. Of a total of 131 DENSE and 134 cine slices, 10 DENSE and 7 cine slices failed inline analysis. Representative end-systolic strain and displacement maps for a cardiotoxicity patient and a heart failure patient—computed via inline DENSE, inline StrainNET, and offline DENSE—are shown in Fig. 1, with corresponding strain-time curves in Fig. 2. Bland–Altman and correlation analyses of whole-slice end-systolic strain (Fig. 3A–D) demonstrated strong agreement for inline DENSE vs. offline DENSE (bias: –0.0%, LOA: –0.7 to 0.7%, r = 0.98) and for StrainNET vs. offline DENSE (bias: 0.4%, LOA: –2.7 to 3.6%, r = 0.80). Segmental comparisons (Fig. 3E–H) showed excellent agreement for inline DENSE (bias: –0.1%, LOA: –4.1 to 3.9%, r = 0.93) and good agreement for StrainNET (bias: 2.2%, LOA: –6.3 to 10.8%, r = 0.71).

Conclusion: Fully automated inline DENSE and StrainNET demonstrated excellent to good agreement, respectively, with offline user-interactive DENSE. Future work will develop automated quality control algorithms for inline analysis and extend validation to broader patient populations.

Fig.1: Example of end systolic strain and displacement maps for a cardiotoxicity patient (row:1-2) and a heart failure patient (row:3-4) for offline user-interactive DENSE, inline DENSE, and inline StrainNET.
Fig. 2: Example of segmental and whole-slice strain-time curves for a heart failure patient for offline user-interactive DENSE, inline DENSE, and inline StrainNET.
Fig. 3: Correlation and Bland-Altman plots for whole-slice and segmental circumferential strain (Ecc) for inline StrainNET and inline DENSE vs offline user-interactive DENSE.