🤖 AI

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

This paper proposes a method that leverages high-resolution cardiac CTA data to train a single neural implicit function, enabling the reconstruction of accurate, smooth, and anatomically plausible cardiac shapes from anisotropic CMRI scans to overcome resolution limitations in cardiac shape analysis.

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

Published 2026-02-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Carolina Brás, Soufiane Ben Haddou, Thijs P. Kuipers, Laura Alvarez-Florez, R. Nils Planken, Fleur V. Y. Tjong, Connie 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

The Big Problem: The "Blurry" Heart Scan

Imagine you are trying to build a 3D model of a heart using a stack of 2D photographs.

The Good News: The photos show the heart's details very clearly from side to side (like looking at a slice of bread).
The Bad News: The photos are taken far apart from each other (like slices of bread with huge gaps in between). Also, the patient might have breathed or moved slightly between shots.

This is what happens with Cardiac MRI (CMRI). It's the standard way doctors look at hearts, but the images are "anisotropic." Think of it like a low-resolution video game where the characters look sharp from the front but are stretched and blocky from the side. When doctors try to measure the heart's shape or volume from these "blocky" slices, the results are often inaccurate or jagged.

The Solution: Borrowing a "High-Def" Blueprint

The researchers asked: "What if we had a perfect, high-resolution 3D blueprint of a heart to help us fix the blurry photos?"

They found that CT Angiography (CTA) scans (a different type of heart scan) are like high-definition 3D models. They are incredibly detailed and show the heart in smooth, continuous 3D. However, CTA scans aren't always available for every patient, and they use radiation, so doctors can't just use them for everything.

The clever idea: Use the high-quality CTA scans to teach a computer what a healthy heart shape looks like, and then use that knowledge to fix the blurry MRI scans.

The Magic Tool: The "Shape Translator" (Neural Implicit Function)

The team built a special AI program called a Neural Implicit Representation (INR). Here is how it works, using a metaphor:

Imagine a master sculptor who has studied thousands of perfect clay hearts (the CTA data). This sculptor has memorized the "rules" of heart shapes—how the walls curve, how the chambers connect, and how the muscle thickness varies.

The Training: The AI "sculptor" looks at the perfect CTA hearts and learns a single, universal language of heart shapes. It learns that the left ventricle, right ventricle, and heart muscle are all connected and move together.
The Translation: When the AI gets a blurry, blocky MRI scan (the "bad" data), it doesn't just try to smooth it out. Instead, it asks: "Based on what I learned from the perfect CTA hearts, what does this blurry mess actually look like in 3D?"
The Result: The AI generates a smooth, high-resolution 3D model that fills in the gaps, corrects the misalignments, and creates a shape that is anatomically correct, even though the original input was low-quality.

Why This Paper is Special

Previous attempts to do this had some limitations:

They were picky: They could only fix one part of the heart at a time (like just the left ventricle).
They needed too much data: They required hundreds of perfect CTA scans to learn.
They were slow: The training took a long time.

This paper's breakthrough:

All-in-One: The AI learns to fix the Left Ventricle, Right Ventricle, and the heart muscle simultaneously. It understands that these parts are neighbors and affect each other.
Efficient: It learned the same (or better) results using three times less data than previous methods.
Fast: It trains in about 1 hour and 40 minutes, making it much closer to being usable in a real hospital.

The Results: From "Blocky" to "Smooth"

The researchers tested this by taking low-quality MRI slices and asking the AI to reconstruct the full 3D heart.

Before: The 3D model looked jagged and disconnected, like a low-poly video game character.
After: The AI produced a smooth, continuous, and realistic heart shape.
Accuracy: When they compared the AI's reconstruction to the "gold standard" high-res scans, the shapes matched very closely (91% accuracy for the right ventricle and 75% for the muscle).

The Bottom Line

Think of this technology as a smart auto-complete feature for heart scans. Just as your phone predicts the rest of a sentence based on the first few words, this AI predicts the rest of the heart's 3D shape based on a few blurry MRI slices, using the "memory" of perfect CTA scans.

This means doctors could eventually get a crystal-clear, 3D understanding of a patient's heart anatomy from standard, quick MRI scans, leading to better diagnoses and treatment plans without needing more radiation or longer scan times.

1. Problem Statement

Cardiovascular Magnetic Resonance Imaging (CMRI) is a gold standard for assessing heart morphology and function. However, standard short-axis (SAX) CMRI scans suffer from anisotropy: they have high in-plane resolution but low through-plane resolution (thick slices with gaps). This leads to:

Suboptimal representation of continuous cardiac shapes.
Misaligned slices and spatial discontinuities.
Inaccurate extraction of clinically relevant metrics (e.g., volumes, strain, ejection fraction) when using standard segmentation methods like Marching Cubes.

While Computed Tomography Angiography (CTA) offers high-resolution, near-isotropic data, it is a different modality. Previous attempts to use CTA to improve CMRI analysis were limited to single anatomical structures (e.g., Left Ventricle Blood Pool only) or required high-resolution CMRI training data, which is rarely available in clinical practice.

2. Methodology

The authors propose a Neural Implicit Representation (INR) framework that leverages high-resolution CTA data to learn a shape prior, enabling the reconstruction of multiple cardiac structures from sparse, anisotropic CMRI data.

A. Data Preparation

Training Data (Set 1): 153 high-resolution CTA scans. Reference segmentations for Left Ventricle Blood Pool (LVBP), Myocardium (MYO), Right Ventricle (RV), and Left Atrium (LA) were generated automatically.
Testing Data (Set 2): 140 paired SAX and 4-chamber (4CH) CMRI scans with lower resolution.
Coordinate Normalization: All coordinates are normalized to a $[-1, 1]^3$ domain and rotated into a shared reference system (aligning the RV-to-LV and LV-to-LA vectors with basis axes) to remove orientation dependency and focus on shape features.

B. Model Architecture: Single Implicit Neural Representation

The core innovation is a single auto-decoder Multilayer Perceptron (MLP) that jointly represents multiple shapes (MYO and RV) using a shared latent space.

Input: A 3D spatial point $\vec{x}$ and a patient-specific multi-shape latent vector $\vec{z}_P$ .
Output: A 2D vector representing the Signed Distance Function (SDF) values for both the Myocardium ( $SDF_{MYO}$ $S D F_{M Y O}$ ) and Right Ventricle ( $SDF_{RV}$ $S D F_{R V}$ ).
- $f_\theta(\vec{x}, \vec{z}_P) \rightarrow [SDF_{MYO}(\vec{x}), SDF_{RV}(\vec{x})]^T$
Architecture: 8 fully connected layers (7 hidden layers of 512 dimensions with ReLU activation, followed by a 2D output layer). Weight normalization is applied to all layers.
Multi-Shape Latent Vector ( $\vec{z}_P$ ): This vector encodes both patient-specific variations and shared anatomical features (e.g., the septal wall) across the different structures.

C. Training and Loss Function

Loss: The model minimizes the Mean Squared Error (MSE) between predicted and reference SDF values.
Regularization: An L2 regularization term ( $\lambda \|\vec{z}_P\|^2$ ) is added to the latent vector to prevent overfitting.
Simplification: Unlike previous DeepSDF-based works, the authors removed the hyperbolic tangent term from the loss function. This simplification, combined with the joint modeling, significantly reduced training time and computational load.
Optimization: The model weights ( $\theta$ ) and latent vectors ( $\vec{z}_P$ ) are jointly optimized during training using AdamW.

D. Inference (Reconstruction)

Latent Vector Optimization: For a new CMRI patient, the model weights are frozen. The latent vector $\vec{z}_P$ is optimized by minimizing the L1-norm difference between the predicted SDF and the SDF derived from the sparse CMRI point cloud.
Surface Extraction: Once $\vec{z}_P$ is found, the 3D shapes are reconstructed by extracting the zero-level sets (iso-surfaces) of the SDFs using the Marching Cubes algorithm.

3. Key Contributions

Joint Multi-Shape Representation: The first method to use a single INR to simultaneously reconstruct the Myocardium (modeling both endocardial and epicardial surfaces) and the Right Ventricle, capturing spatial dependencies between structures.
Cross-Modality Learning: Successfully transfers shape priors from high-resolution CTA to low-resolution, anisotropic CMRI without requiring high-resolution CMRI training data.
Efficiency and Scalability:
- Achieves comparable performance to previous state-of-the-art methods (Sander et al.) while using 3x less CTA training data (100 patients vs. 341).
- Reduces training time from 6 hours to **1 hour 40 minutes** via a simplified loss function.
- Inference is rapid (~4 seconds per patient).
Robustness to Anisotropy: Demonstrates the ability to reconstruct smooth, anatomically plausible 3D shapes from sparse point clouds derived from SAX and 4CH CMRI slices.

4. Results

The method was evaluated on 4-chamber (4CH) slices extracted from the reconstructed 3D shapes and compared against reference high-resolution 4CH masks.

Quantitative Performance (Dice Similarity Coefficient - DSC):
- RV: Proposed method achieved 0.91 ± 0.07 (vs. 0.81 for direct SAX extraction).
- MYO: Proposed method achieved 0.75 ± 0.13 (vs. 0.57 for direct SAX extraction).
- LVBP: Achieved 0.93 ± 0.05.
Geometric Accuracy (Hausdorff Distance - HD):
- RV: Reduced HD from 10.15 mm (SAX) to 6.21 mm.
- MYO: Reduced HD from 12.99 mm (SAX) to 7.53 mm.
Generalization Test: The model was tested on simulated anisotropic point clouds derived from CTA data. The reconstruction quality using sparse (SAX-like) CTA points was only marginally lower than using dense (isotropic) CTA points, proving the model's ability to generalize from low-resolution inputs.

5. Significance and Conclusion

This work addresses a critical bottleneck in cardiac imaging: the inability to perform accurate 3D shape analysis on standard, anisotropic CMRI scans. By leveraging a single implicit neural network trained on CTA data, the authors provide a robust solution that:

Corrects Anisotropy: Fills in missing through-plane information to create smooth, continuous 3D cardiac models.
Enables Clinical Utility: The speed of inference and the accuracy of the reconstructed shapes make the method viable for clinical integration, potentially improving the calculation of ejection fraction and strain analysis.
Facilitates Research: The code is publicly available, and the method opens new avenues for studying cardiac morphology and function using high-resolution 3D representations derived from routine clinical scans.

The authors note that while the method performs well, regions with very low CMRI coverage (e.g., the apex) still present challenges, suggesting future work should incorporate additional long-axis views (2-chamber, 3-chamber) to further refine the shape priors.