Oral Abstract 1: Validating DENSE strain using synthetic data from a modular simulation

Presenting Author(s)

• HB

  Hugo Barbaroux, MSc

  PhD Student
  King's College London, United Kingdom

Primary Author(s)

• HB

  Hugo Barbaroux, MSc

  PhD Student
  King's College London, United Kingdom

Co-Author(s)

• HB

  Hugo Barbaroux, MSc

  PhD Student
  King's College London, United Kingdom
• ML

  Michael Loecher, PhD

  Research Scientist
  Stanford University
• KK

  Karl P. Kunze, PhD

  Senior Cardiac MR Scientist
  Siemens Healthineers, United Kingdom
• 

  Daniel B. Ennis, PhD

  Professor
  Stanford University
• 

  Sonia Nielles-Vallespin, PhD, MSc, BSc

  Senior Lecturer
  Imperial College London
• 

  Andrew D. Scott, PhD, FSCMR

  Senior Lecturer
  Imperial College London and Royal Brompton Hospital, United Kingdom
• 

  Alistair A. Young, PhD

  Professor
  King's College London, United Kingdom

Background: Cine Displacement Encoding with Stimulated Echoes (DENSE) is a CMR technique providing myocardial displacement and strain at a pixel level [1]. Deep learning methods enable automated segmentation and strain estimation [2, 3], but lack validation with a true ground-truth strain, as most techniques only focus on replicating manual results, which may be biased [4]. We present here a modular pipeline for generating realistic synthetic DENSE images incorporating ground-truth strain data, and using it to evaluate DENSEanalysis, a widely used tool for DENSE processing [5].

Methods: The pipeline consists of three modules: anatomy generation, cardiac deformation, and DENSE simulation. The anatomy generation module includes background tissues (optional) and an end-diastolic cardiac shape for which three methods provide variability: XCAT [6] phantom, in vivo cardiac segmentations, and annular phantoms. The cardiac deformation module uses spatiotemporal polynomial deformation fields constrained to contraction-like motion at high resolution [7]. The DENSE module is based on Bloch simulations, considering each point of the high-resolution displacement field as spins mapped to MRI values (M0, T1, T2), leading to a lower-resolution DENSE image. Spiral sampling with four interleaves was simulated, with cine sequences from 25 to 75 frames and optional frame interpolation to replicate in vivo acquisitions. 180 short-axis cases at three different SNRs (5, 10, and 15, typical of in vivo dataset) were generated and evaluated using DENSEanalysis with a range of spatial regularization values from 0.9 (typical) [8] to 0.1 (negligible).

Results: Fig.1 shows an example of magnitude and phase DENSE images at SNR=15 compared to an in vivo example. Over all cases, the average pixelwise difference between the ground-truth strain and the DENSE strain for the radial component was 0.06±0.08 at 0.9 regularization, 0.05±0.11 at 0.6, 0.03±0.16 at 0.3, and 0.02±0.24 at 0.1. The circumferential strain difference was -0.01±0.02 at 0.9, -0.02±0.04 at 0.6, -0.02±0.07 at 0.3, and -0.03±0.11 at 0.1. This shows that typical regularization parameters underestimate the radial strain, even when compensating well for the noise. There is a bias towards underestimating the circumferential strain which is more severe for decreasing regularization, likely due to increased effects of noise. Regularization has a high impact on estimating strain, and optimal parameters vary depending on the target strain component (Fig.2 & 3).

Conclusion: Careful validation is necessary for establishing the clinical use of DENSE-derived strain as a biomarker, which can be achieved through image simulation. In DENSEanalysis, to balance segmental bias and pixel variability, an optimal regularization value might be (0.3 – 0.6) for the radial strain depending on the SNR, and 0.9 for the circumferential strain. This simulation pipeline potentially provides accurate ground-truth data for deep-learning strain estimation algorithms.

Figure 1: Comparison example of in-vivo (bottom row) and simulated (top row, without background tissues, high SNR) end-systolic (ES) DENSE images (magnitude and phases).

Figure 2: Pixelwise circumferential strain errors - the effects of noise and spatial regularization. Circumferential strain metrics of DENSE strain extracted with DENSEanalysis at different regularization parameters (0.9, 0.6, 0.3, 0.1) and SNR (low=5, mid=10, high=15) against ground-truth strain. The top row (a) displays the average segmental pixel strain difference between the ground truth and the prediction (negative values indicate an under-estimation of the strain), for different transmural segments at different SNR levels. Bottom row (b) shows pixel standard deviations from ground truth.

Figure 3: Pixelwise radial strain errors - the effects of noise and spatial regularization. Radial strain metrics of DENSE strain extracted with DENSEanalysis at different regularization parameters (0.9, 0.6, 0.3, 0.1) and SNR (low=5, mid=10, high=15) against ground-truth strain. The top row (a) displays the average segmental pixel strain difference between the ground truth and the prediction (positive values indicate an under-estimation of the strain), for different transmural segments at different SNR levels. Bottom row (b) shows pixel standard deviations from ground truth.