Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Synthesis of Late Gadolinium Enhancement Images via Implicit Neural Representations for Cardiac Scar Segmentation

Soufiane Ben Haddou, Laura Alvarez-Florez, Erik J. Bekkers, Fleur V. Y. Tjong, Ahmad S. Amin, Connie R. Bezzina, Ivana Išgum

TL;DR

The paper addresses data scarcity in LGE cardiac scar segmentation by proposing a unified INR-diffusion framework that jointly models LGE images and myocardium/fibrosis masks, compresses these representations into a latent space, and synthesizes new image–mask pairs for augmentation. Using 133 patient scans (105 for training INR, 28 for testing), the method demonstrates that adding synthetic volumes improves fibrosis segmentation Dice from $0.509 \pm 0.08$ to $0.524 \pm 0.07$, with the strongest gains in apical regions. The approach combines Sinusoidal Representation Networks (SIREN) for joint image-mask encoding, INR2VEC for latent compression to $512$ dimensions, and Elucidated Diffusion Models to generate diverse, anatomically plausible samples, evaluated via nnU-Net. This annotation-free data generation strategy mitigates data scarcity and supports more robust automated scar quantification in LGE MRI.

Abstract

Late gadolinium enhancement (LGE) imaging is the clinical standard for myocardial scar assessment, but limited annotated datasets hinder the development of automated segmentation methods. We propose a novel framework that synthesises both LGE images and their corresponding segmentation masks using implicit neural representations (INRs) combined with denoising diffusion models. Our approach first trains INRs to capture continuous spatial representations of LGE data and associated myocardium and fibrosis masks. These INRs are then compressed into compact latent embeddings, preserving essential anatomical information. A diffusion model operates on this latent space to generate new representations, which are decoded into synthetic LGE images with anatomically consistent segmentation masks. Experiments on 133 cardiac MRI scans suggest that augmenting training data with 200 synthetic volumes contributes to improved fibrosis segmentation performance, with the Dice score showing an increase from 0.509 to 0.524. Our approach provides an annotation-free method to help mitigate data scarcity.The code for this research is publicly available.

Synthesis of Late Gadolinium Enhancement Images via Implicit Neural Representations for Cardiac Scar Segmentation

TL;DR

The paper addresses data scarcity in LGE cardiac scar segmentation by proposing a unified INR-diffusion framework that jointly models LGE images and myocardium/fibrosis masks, compresses these representations into a latent space, and synthesizes new image–mask pairs for augmentation. Using 133 patient scans (105 for training INR, 28 for testing), the method demonstrates that adding synthetic volumes improves fibrosis segmentation Dice from to , with the strongest gains in apical regions. The approach combines Sinusoidal Representation Networks (SIREN) for joint image-mask encoding, INR2VEC for latent compression to dimensions, and Elucidated Diffusion Models to generate diverse, anatomically plausible samples, evaluated via nnU-Net. This annotation-free data generation strategy mitigates data scarcity and supports more robust automated scar quantification in LGE MRI.

Abstract

Late gadolinium enhancement (LGE) imaging is the clinical standard for myocardial scar assessment, but limited annotated datasets hinder the development of automated segmentation methods. We propose a novel framework that synthesises both LGE images and their corresponding segmentation masks using implicit neural representations (INRs) combined with denoising diffusion models. Our approach first trains INRs to capture continuous spatial representations of LGE data and associated myocardium and fibrosis masks. These INRs are then compressed into compact latent embeddings, preserving essential anatomical information. A diffusion model operates on this latent space to generate new representations, which are decoded into synthetic LGE images with anatomically consistent segmentation masks. Experiments on 133 cardiac MRI scans suggest that augmenting training data with 200 synthetic volumes contributes to improved fibrosis segmentation performance, with the Dice score showing an increase from 0.509 to 0.524. Our approach provides an annotation-free method to help mitigate data scarcity.The code for this research is publicly available.
Paper Structure (14 sections, 3 equations, 4 figures, 1 table)

This paper contains 14 sections, 3 equations, 4 figures, 1 table.

Table of Contents

  1. INTRODUCTION
  2. Dataset
  3. METHODS & EXPERIMENTS
  4. Joint Image-Mask Implicit Neural Representation
  5. Latent Space Compression and Generation
  6. Diffusion-based Synthesis
  7. Image segmentation
  8. Experimental Setup
  9. RESULTS
  10. Evaluation of Synthesized LGE Images
  11. Volumetric and Geometric Consistency
  12. Analysis of the Automatic Segmentations
  13. DISCUSSION
  14. CONCLUSION

Figures (4)

  • Figure 1: Overview of our proposed framework. An INR model captures both LGE image intensities and its segmentation masks, which are then compressed into a latent space. A diffusion model operates on this latent space to generate synthetic LGE data with anatomically consistent segmentation masks.
  • Figure 2: Comparison of real (top) and an example synthetic (bottom) LGE short-axis slices at basal (slice 3), mid-ventricular (slice 6), and apical (slice 9) levels. Top: Slices with ground truth (a) or generated (b) segmentation contours (red = myocardium, blue = fibrosis). Bottom: Same slices without overlay showing image quality and anatomical consistency.
  • Figure 3: PSNR and SSIM for synthetic vs. real samples, evaluated on the middle LGE slice from each volume.
  • Figure 4: Comparison of myocardium and fibrosis mask geometries. (a) Real geometry displayed as a grey mesh, with myocardium in blue and fibrosis in red. (b) Synthetic geometry shown using the same representation.