🤖 AI

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

This paper proposes a novel framework that combines implicit neural representations with denoising diffusion models to synthesize anatomically consistent late gadolinium enhancement images and segmentation masks, thereby improving cardiac fibrosis segmentation performance by mitigating data scarcity through annotation-free data augmentation.

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

Published 2026-02-13

📖 5 min read🧠 Deep dive

CC BY 4.0

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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to spot a specific type of scar on a human heart. This isn't a regular scar; it's a hidden injury inside the heart muscle that doctors can only see using a special, high-tech camera called an LGE MRI.

Here is the problem: To teach the robot, you need thousands of examples. But getting these examples is a nightmare.

The Images are fuzzy: The heart is always moving, and the images are often blurry or have weird shadows.
The Labels are rare: To teach the robot, a human expert has to manually trace every single scar on every single slice of the heart. This takes hours, costs a fortune, and there just aren't enough experts to do it for thousands of patients.

This paper presents a clever solution: Instead of waiting for more human experts, we teach the computer to "dream" up new, realistic heart scans and their corresponding scar maps.

Here is how they did it, broken down into simple analogies:

1. The "Infinite Zoom" Map (Implicit Neural Representations)

Usually, computers store images like a grid of pixels (like a mosaic). If you zoom in too far, it gets blocky.

The researchers used a technique called Implicit Neural Representations (INRs). Think of this not as a grid of pixels, but as a mathematical recipe or a GPS coordinate system.

Instead of storing a picture, the computer learns a formula that says: "If you are at coordinate X, Y, and Z, the heart muscle here is bright, and the scar here is dark."
Because it's a formula, you can zoom in infinitely without losing quality. It captures the smooth, continuous shape of the heart perfectly.

2. The "Compressed Suitcase" (Latent Space)

These mathematical recipes are huge and complicated. You can't easily shuffle them around to make new ones.

So, the researchers built a "compressor" (called INR2VEC). Imagine taking a giant, messy suitcase full of clothes (the complex heart data) and folding it down into a tiny, neat suitcase that fits in your pocket.

This tiny suitcase contains all the essential information about the heart's shape and where the scars are, but it's small enough to be easily manipulated.

3. The "Dreaming Artist" (Diffusion Models)

Now comes the magic. They took a Diffusion Model (a type of AI famous for generating art) and taught it to play with these tiny suitcases.

How it works: Imagine the AI starts with a bag of static noise (like TV snow). It slowly "denoises" it, step by step, guided by the patterns it learned from the real hearts.
The Result: The AI "dreams" up a brand new, tiny suitcase. This suitcase represents a heart that never existed before, but it looks and feels exactly like a real human heart.
The Best Part: Because the AI was trained on the "recipe" (the INR), it doesn't just make a picture; it makes the picture and the perfect map of the scars at the same time. It's like a painter who paints a landscape and simultaneously draws the exact outline of the trees and rocks on a separate piece of paper, perfectly aligned.

4. The "Practice Test" (Segmentation)

To see if this fake data actually helps, they took a standard robot (an AI called nnU-Net) and gave it two training sets:

Group A: Only real, human-labeled heart scans (105 patients).
Group B: The same 105 real scans, plus 200 "dreamed up" synthetic scans.

The Result:

The robot in Group B got much better at finding the scars.
Specifically, it got 9.5% better at finding scars in the very tip (apex) of the heart. This is the hardest area to see because scars there are tiny and rare.
The robot didn't get confused by the fake data; instead, the fake data taught it what a "rare scar" looks like, making it a sharper detective.

Why This Matters

Think of this like a medical student.

Before: The student only gets to study 100 real patients. They miss the tricky cases because they haven't seen enough variety.
After: The student studies those 100 real patients, plus 200 "simulated" patients generated by a super-smart computer. These simulated patients show every possible variation of heart scars.
The Outcome: When the student faces a real patient with a tricky, rare scar, they recognize it immediately because they've "seen" it before in the simulation.

The Bottom Line

This paper introduces a way to create infinite, perfect training data without needing a single extra human to draw a line. It solves the "data scarcity" problem by using AI to dream up realistic heart scans and their scar maps, helping future AI doctors diagnose heart disease more accurately, especially in the hardest-to-see areas.

1. Problem Statement

Late Gadolinium Enhancement (LGE) cardiac MRI is the clinical standard for assessing myocardial scars (fibrosis), which is critical for patient prognosis and treatment planning. However, the development of automated segmentation algorithms is severely hindered by data scarcity:

Annotation Difficulty: Manual delineation of scars requires specialized expertise, is labor-intensive, and suffers from inter- and intra-observer variability.
Image Quality: LGE images often have low spatial resolution, motion artifacts, and variable contrast, making scar boundaries (which are often diffuse and heterogeneous) difficult to define.
Limitations of Existing Augmentation: Traditional augmentation techniques fail to capture complex spatial relationships. Domain translation methods (e.g., converting cine CMR to LGE) often still require manual annotations of the generated images or are limited by the availability of source data.

The core challenge is generating synthetic LGE images paired with accurate segmentation masks without requiring manual annotation, thereby augmenting training datasets to improve deep learning performance.

2. Methodology

The authors propose a unified framework combining Implicit Neural Representations (INRs) and Denoising Diffusion Models to jointly synthesize LGE images and their corresponding segmentation masks.

A. Joint Image-Mask Representation (INR)

Instead of treating images and masks as discrete 2D/3D grids, the method models them as continuous functions using Sinusoidal Representation Networks (SIREN).

Input: 3D spatial coordinates $x = [x, y, z]^T$ .
Output: A multi-task network $F_\theta(x)$ $F_{θ} (x)$ that simultaneously predicts:
- Grayscale image intensity $\hat{I}(x)$ .
- Binary mask for myocardium $\hat{M}_1(x)$ .
- Binary mask for fibrosis $\hat{M}_2(x)$ .
Architecture: A shared backbone $\phi_{shared}$ extracts common anatomical features, followed by task-specific heads $\psi_{task}$ .
Loss Function: A composite loss minimizes $L_2$ norm for image reconstruction and Binary Cross-Entropy (BCE) for mask prediction, ensuring anatomical consistency between the image and its segmentation.

B. Latent Space Compression (INR2VEC)

Directly manipulating INR parameters is computationally expensive due to high dimensionality (~260K parameters).

The authors employ INR2VEC to compress the flattened parameter vector $\theta$ of each trained INR into a compact 512-dimensional latent vector $z$ .
This encoder preserves essential anatomical information while enabling efficient data manipulation.

C. Diffusion-Based Synthesis

A generative model operates within this compressed latent space to create new data.

Model: An Elucidated Diffusion Model (EDM) is used.
Process:
1. Forward: Gaussian noise is progressively added to latent vectors over 1000 timesteps.
2. Reverse: A neural network learns to denoise, sampling from a Gaussian distribution to generate new latent vectors $z_0$ .
3. Decoding: The new latent vector is decoded back into INR parameters via the INR2VEC decoder.
4. Generation: The decoded INR is queried at spatial coordinates to generate the final synthetic LGE image and segmentation masks.

D. Evaluation Setup

Dataset: 133 patients (105 for training, 28 for testing) with short-axis LGE CMR scans.
Segmentation Model: nnU-Net (state-of-the-art) was trained to segment myocardium and fibrosis.
Experiment: Three models were compared:
1. Baseline (105 real volumes).
2. Augmented +100 synthetic volumes.
3. Augmented +200 synthetic volumes.

3. Key Contributions

Annotation-Free Synthesis: The first method to utilize latent diffusion for synthesizing paired LGE images and segmentation masks without requiring manual annotation of the synthetic data.
Anatomical Consistency: By modeling images and masks jointly within a single INR framework, the method guarantees that generated scars are anatomically plausible and spatially consistent with the myocardium, avoiding the misalignment issues common in separate generation pipelines.
Continuous Representation: Utilizing INRs allows the capture of fine-grained spatial relationships and continuous geometry, overcoming the limitations of discrete voxel-based augmentation.
Data Scarcity Mitigation: Demonstrates a viable pathway to expand small medical datasets effectively for deep learning training.

4. Results

The study evaluated both the quality of the synthesized data and its impact on downstream segmentation tasks.

A. Quality of Synthesized Data

Visual Quality: Synthetic slices preserved global cardiac geometry, consistent wall thickness, and coherent myocardial contours. Fibrosis regions appeared at anatomically plausible locations within the myocardial wall.
Metrics: While PSNR and SSIM values were moderate (indicating the synthetic data does not replicate specific real scans voxel-for-voxel), the structural similarity was highest within the myocardial and fibrosis regions, confirming the model focuses on clinically relevant areas.
3D Coherence: The method maintained volumetric coherence across slices, avoiding the "slice-to-slice" inconsistencies often seen in 2D generation methods.

B. Segmentation Performance

Augmenting the training set with synthetic data yielded specific improvements:

Fibrosis Segmentation: The Dice score improved from 0.509 (baseline) to 0.524 (+2.9%) with 200 synthetic volumes.
- Apical Slices: The most significant gain was observed in apical slices (+9.5%), where fibrosis is typically small, heterogeneous, and underrepresented in real data.
Myocardium Segmentation: Performance remained relatively stable (Dice ~0.815), with a negligible decrease when adding more synthetic data. This suggests the real dataset already sufficiently covers myocardial anatomy, and synthetic data primarily benefits the scar (fibrosis) task.

5. Significance and Conclusion

This research addresses a critical bottleneck in cardiac AI: the lack of annotated data for scar segmentation.

Clinical Impact: By improving fibrosis segmentation, particularly in difficult apical regions where data is scarce, the method supports more accurate patient risk stratification and treatment planning.
Methodological Shift: The approach shifts the paradigm from "pixel-perfect reconstruction" to "anatomically coherent generation." The goal is not to mimic a specific patient's scan exactly, but to generate diverse, structurally valid training samples that teach the segmentation network to recognize scar patterns.
Future Directions: The authors suggest that while myocardial segmentation is robust, future work could focus on enhancing boundary sharpness (e.g., via adversarial training) and investigating why synthetic data impacts fibrosis and myocardium segmentation differently.

In summary, the proposed INR-Diffusion framework successfully generates high-quality, anatomically consistent synthetic LGE data that significantly boosts the performance of automated scar segmentation models, offering a scalable solution to data scarcity in cardiac imaging.