Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fighting MRI Anisotropy: Learning Multiple Cardiac Shapes From a Single Implicit Neural Representation

Carolina Brás, Soufiane Ben Haddou, Thijs P. Kuipers, Laura Alvarez-Florez, R. Nils Planken, Fleur V. Y. Tjong, Connie Bezzina, Ivana Išgum

TL;DR

The paper tackles MRI anisotropy in cardiac shape analysis by leveraging high-resolution CTA data to train a single implicit neural representation that jointly models the left-ventricular blood pool, myocardium, and right ventricle as zero level sets of signed distance functions. A multi-shape latent space, learned via an auto-decoder MLP, encodes patient-specific anatomy and enables high-resolution, smooth reconstructions from SAX CMRI segmentations, with surfaces extracted through Marching Cubes. The approach demonstrates strong 4CH reconstruction performance and generalizes across imaging modalities, achieving comparable results with isotropic CTA points and SAX-like sampling while reducing CTA data requirements and training time. This method promises improved cardiac shape analysis from common CMRI protocols by incorporating CTA-derived shape priors for high-fidelity, multi-structure reconstructions.

Abstract

The anisotropic nature of short-axis (SAX) cardiovascular magnetic resonance imaging (CMRI) limits cardiac shape analysis. To address this, we propose to leverage near-isotropic, higher resolution computed tomography angiography (CTA) data of the heart. We use this data to train a single neural implicit function to jointly represent cardiac shapes from CMRI at any resolution. We evaluate the method for the reconstruction of right ventricle (RV) and myocardium (MYO), where MYO simultaneously models endocardial and epicardial left-ventricle surfaces. Since high-resolution SAX reference segmentations are unavailable, we evaluate performance by extracting a 4-chamber (4CH) slice of RV and MYO from their reconstructed shapes. When compared with the reference 4CH segmentation masks from CMRI, our method achieved a Dice similarity coefficient of 0.91 $\pm$ 0.07 and 0.75 $\pm$ 0.13, and a Hausdorff distance of 6.21 $\pm$ 3.97 mm and 7.53 $\pm$ 5.13 mm for RV and MYO, respectively. Quantitative and qualitative assessment demonstrate the model's ability to reconstruct accurate, smooth and anatomically plausible shapes, supporting improvements in cardiac shape analysis.

Fighting MRI Anisotropy: Learning Multiple Cardiac Shapes From a Single Implicit Neural Representation

TL;DR

The paper tackles MRI anisotropy in cardiac shape analysis by leveraging high-resolution CTA data to train a single implicit neural representation that jointly models the left-ventricular blood pool, myocardium, and right ventricle as zero level sets of signed distance functions. A multi-shape latent space, learned via an auto-decoder MLP, encodes patient-specific anatomy and enables high-resolution, smooth reconstructions from SAX CMRI segmentations, with surfaces extracted through Marching Cubes. The approach demonstrates strong 4CH reconstruction performance and generalizes across imaging modalities, achieving comparable results with isotropic CTA points and SAX-like sampling while reducing CTA data requirements and training time. This method promises improved cardiac shape analysis from common CMRI protocols by incorporating CTA-derived shape priors for high-fidelity, multi-structure reconstructions.

Abstract

The anisotropic nature of short-axis (SAX) cardiovascular magnetic resonance imaging (CMRI) limits cardiac shape analysis. To address this, we propose to leverage near-isotropic, higher resolution computed tomography angiography (CTA) data of the heart. We use this data to train a single neural implicit function to jointly represent cardiac shapes from CMRI at any resolution. We evaluate the method for the reconstruction of right ventricle (RV) and myocardium (MYO), where MYO simultaneously models endocardial and epicardial left-ventricle surfaces. Since high-resolution SAX reference segmentations are unavailable, we evaluate performance by extracting a 4-chamber (4CH) slice of RV and MYO from their reconstructed shapes. When compared with the reference 4CH segmentation masks from CMRI, our method achieved a Dice similarity coefficient of 0.91 0.07 and 0.75 0.13, and a Hausdorff distance of 6.21 3.97 mm and 7.53 5.13 mm for RV and MYO, respectively. Quantitative and qualitative assessment demonstrate the model's ability to reconstruct accurate, smooth and anatomically plausible shapes, supporting improvements in cardiac shape analysis.
Paper Structure (11 sections, 3 equations, 5 figures, 2 tables)

This paper contains 11 sections, 3 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Data
  3. Method
  4. Data Sampling and Preparation
  5. Implicit multi-shape learning
  6. Multi-shape reconstruction and completion
  7. Experiments and Results
  8. Auto-decoder Training
  9. Shape Reconstruction and Evaluation
  10. Evaluation of Learned CTA Multi-shape Prior
  11. Discussion and Conclusion

Figures (5)

  • Figure 1: Training with point clouds from CTA segmentations. During training, the model weights and multi-shape latent vectors are jointly optimized. Each anatomical shape is then implicitly represented by the zero iso-surface of its SDF. LV and RV are represented in purple and pink, respectively.
  • Figure 2: Inference with point clouds from SAX and 4-chamber CMRI segmentations. During inference, the model weights are fixed and the latent vectors are firstly optimized for each multi-shape point cloud. LV and RV are represented in purple and pink, respectively.
  • Figure 3: Examples illustrating a 4CH CMRI overlaid with reference, reconstructed and SAX-based segmentations of LVBP (darker purple), MYO (lighter purple) and RV (pink).
  • Figure 4: Example illustrating each reconstructed shape and the corresponding reference segmentation coordinates. The shapes are rotated around the through-plane axis (0º, 90º and 180º). LV and RV are represented in purple and pink, respectively. Darker purple highlights LVBP, whereas lighter purple depicts MYO. Points in cyan and yellow belong to SAX and 4CH segmentations, respectively. These reconstructions correspond to Example 1 shown in Figure \ref{['fig:cross_sections']}.
  • Figure 5: Example illustrating reconstructed shapes overlaid with their reference CTA point cloud. The shapes are rotated around the through-plane axis (0º, 90º and 180º). LV and RV are represented in purple and pink, respectively. Darker purple highlights LVBP, whereas lighter purple depicts MYO. Points in cyan and magenta belong to MYO and RV surface point clouds, respectively. (a) Reference point cloud ($\mathcal{P}_{\text{Reference}}$) and surface meshes directly extracted from CTA segmentations. (b) Reconstructed surface meshes from a highly isotropic CTA point cloud ($\mathcal{P}_{\text{CTA-ISO}}$). (c) Reconstructed surface meshes from an anisotropic SAX-like simulated CTA point cloud ($\mathcal{P}_{\text{CTA-SAX}}$).