🤖 AI

Neural Implicit 3D Cardiac Shape Reconstruction from Sparse CT Angiography Slices Mimicking 2D Transthoracic Echocardiography Views

This paper proposes a neural implicit function-based method that reconstructs accurate 3D cardiac shapes from sparse CT angiography planes mimicking 2D echocardiography views, demonstrating superior volumetric quantification of heart chambers compared to the clinical standard Simpson's biplane rule.

Original authors: Gino E. Jansen, Carolina Brás, R. Nils Planken, Mark J. Schuuring, Berto J. Bouma, Ivana Išgum

Published 2026-04-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Gino E. Jansen, Carolina Brás, R. Nils Planken, Mark J. Schuuring, Berto J. Bouma, 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 trying to figure out the exact shape of a complex, squishy object (like a heart) just by looking at a few thin slices of it, kind of like trying to guess the shape of a whole loaf of bread by only looking at three or four slices of toast.

That is essentially the challenge doctors face with 2D Echocardiography (ultrasound of the heart). They get clear images of the heart from a few specific angles, but they have to guess the 3D shape and volume based on those flat pictures. Usually, they use a math formula called "Simpson's rule" to do this guesswork. But if the ultrasound probe isn't perfectly aligned, the guess can be way off, leading to underestimating how big the heart chambers are.

This paper introduces a new, smarter way to solve this puzzle using Artificial Intelligence (AI). Here is how it works, broken down into simple concepts:

1. The "Mental Blueprint" (The Shape Prior)

First, the researchers taught a computer brain (a neural network) what a healthy human heart actually looks like in 3D. They didn't use fake computer hearts; they used high-resolution 3D scans from CT Angiography (a super-detailed X-ray) of real patients.

Think of this like a master sculptor who has studied thousands of real hearts. They have memorized the "rules" of heart shapes: how the left ventricle curves, how the atrium connects, and how the muscle walls thicken. This knowledge is the Shape Prior. The AI now knows, "If I see a slice that looks like this, the rest of the heart probably looks like that."

2. The "Magic Reconstructor" (Neural Implicit Function)

Instead of trying to build the heart pixel-by-pixel, the AI uses a Neural Implicit Function.

The Analogy: Imagine a magical 3D printer that doesn't need a blueprint file. Instead, you give it a tiny "secret code" (a latent vector) and a set of coordinates (x, y, z). The printer instantly knows whether that specific point in space is "inside the heart" or "outside the heart."
By feeding it different secret codes, the printer can create millions of slightly different, realistic heart shapes.

3. The "Jigsaw Puzzle" (Test-Time Optimization)

This is the clever part. When the AI gets a new patient's data, it only has a few 2D slices (mimicking the ultrasound views). It doesn't know the exact 3D shape of this specific patient, nor does it know exactly how the ultrasound probe was held (the angle might be slightly off).

So, the AI plays a game of Jigsaw Puzzle:

Guess the Shape: It picks a "secret code" to generate a 3D heart.
Guess the Angle: It guesses how the 2D slices were rotated in 3D space.
Check the Fit: It compares its generated 3D heart against the actual 2D slices the patient has. Do the lines match?
Adjust and Repeat: If the lines don't match, the AI tweaks the secret code (to change the shape) and the angles (to fix the probe position) and tries again. It does this thousands of times in seconds until the 3D heart fits the 2D slices perfectly.

4. The Results: Why It Matters

The researchers tested this on 40 patients and compared it to the old standard (Simpson's rule).

The Old Way: Like guessing the volume of a water balloon by looking at two flat shadows. It was okay, but often made big mistakes (e.g., missing 8 mL of volume in the left ventricle).
The New Way: Like using the AI to reconstruct the whole balloon from those shadows. It was much more accurate (only missing about 4.8 mL).
The Big Win: For the Left Atrium (a smaller chamber), the old method was wildly inaccurate (missing nearly 38 mL!), while the new AI method was very close (missing only 6.4 mL).

The Bottom Line

This paper proposes a way to turn standard, 2D ultrasound images into highly accurate 3D models of the heart. By combining a deep understanding of what hearts should look like (learned from CT scans) with a smart algorithm that figures out the exact angle of the ultrasound probe, doctors can get a much clearer, more accurate picture of a patient's heart health without needing expensive 3D ultrasound machines.

It's like upgrading from a blurry sketch to a high-definition 3D hologram, just by using a few flat photos and a very smart computer.

1. Problem Statement

Accurate 3D quantification of cardiac anatomy and function is crucial for clinical assessment. While 3D Transthoracic Echocardiography (TTE) exists, 2D TTE remains the clinical standard due to better spatiotemporal resolution. However, 2D TTE relies on geometric assumptions (e.g., Simpson's biplane rule) to estimate volumes from sparse planes. These assumptions are prone to systematic errors, particularly when the ultrasound probe is misaligned with the heart's apex, leading to underestimation of chamber volumes.

Existing deep learning approaches for 3D reconstruction from 2D views often suffer from two main limitations:

Reliance on Synthetic Priors: Many methods use synthetic shape datasets, which may not capture the anatomical variability of real patients.
Catastrophic Forgetting: Test-time optimization methods that update all model weights risk forgetting the learned shape priors.

The authors propose a method to reconstruct complete 3D cardiac shapes from sparse 2D slices that mimic standard apical TTE views, using real patient CT Angiography (CTA) data as the ground truth prior.

2. Methodology

The proposed framework extends the implicit shape reconstruction approach by Amiranashvili et al. [9] with specific adaptations for multi-class cardiac structures and view-pose optimization.

A. Data Preparation

Source: 452 patients with acute ischemic stroke underwent cardiac CTA.
Segmentation: A pre-trained deep learning model generated 3D segmentations for five structures: Left Atrium (LA), Left Ventricle (LV), Right Atrium (RA), Right Ventricle (RV), and LV Myocardium (LV-myo).
Simulation: From the 3D CTA volumes, the authors simulated four standard apical TTE views (A2C, A3C, A4C, A5C). To mimic the uncertainty of free-hand ultrasound acquisition, they applied Gaussian perturbations ( $\sigma = 5.0$ mm) to the anatomical landmarks (apex and chamber centers of mass) used to define the slice planes.

B. Neural Implicit Function (Training Phase)

Architecture: A Multi-Layer Perceptron (MLP) with 8 hidden layers (width 128) and skip connections.
Input: 3D Cartesian coordinates ( $x \in \mathbb{R}^3$ ) concatenated with a learnable latent code ( $z \in \mathbb{R}^{128}$ ) unique to each subject.
Output: A softmax probability vector ( $p \in [0, 1]^6$ ) representing occupancy for 6 classes (background + 5 cardiac structures).
Loss Function: A combination of categorical cross-entropy and multi-class soft Dice loss, plus an $L_2$ regularization term on the latent codes.
Goal: The MLP learns a continuous shape prior distribution of the heart from dense 3D CTA segmentations.

C. Test-Time Optimization (Inference Phase)

At inference, the MLP weights are frozen. The network reconstructs the 3D shape from sparse 2D masks by optimizing two sets of parameters:

Latent Code ( $z$ ): Represents the specific shape of the patient's heart.
Rigid Transformations ( $T_{view}$ ): Parameters (axis-angle rotation $\alpha$ and translation $t$ ) for each of the four input planes.

Optimization Strategy:

Phase 1 (100 steps): Optimize only the latent code $z$ while keeping view poses fixed.
Phase 2 (1000 steps): Jointly optimize $z$ and the view poses ( $\alpha, t$ ).
Objective: Minimize the cross-entropy and Dice loss between the network's predicted 3D shape (rendered to 2D) and the observed sparse segmentation masks.
Constraint: The A4C view is anchored to its true (perturbed) orientation to prevent global drift of the reconstructed volume, allowing direct metric calculation against the reference.

3. Key Contributions

Real-World Shape Priors: Unlike previous works using synthetic data, this method learns shape priors directly from real, high-resolution CTA segmentations of 100 patients, capturing true anatomical variability.
Joint Pose and Shape Optimization: The method introduces a test-time optimization scheme that simultaneously refines the latent shape representation and the rigid view poses. This addresses the critical issue of misalignment in 2D acquisitions without requiring model weight updates (avoiding catastrophic forgetting).
Multi-Class Reconstruction: The framework successfully reconstructs not just the Left Ventricle, but a multi-class whole-heart model including atria, ventricles, and myocardium.
Superiority over Clinical Standards: The method is rigorously benchmarked against Simpson's biplane rule, the current clinical gold standard for 2D TTE volume estimation.

4. Results

The method was evaluated on a held-out test set of 40 patients.

Overall Accuracy: The approach achieved an average Dice coefficient of 0.86 ± 0.04 across all structures.
Volume Estimation (Left Ventricle):
- Proposed Method: 4.88 ± 4.26 mL Mean Absolute Error (MAE).
- Simpson's Rule: 8.14 ± 6.04 mL MAE.
- Result: The proposed method reduced volume error by nearly 40% compared to the clinical standard.
Volume Estimation (Left Atrium):
- Proposed Method: 6.40 ± 7.37 mL MAE.
- Simpson's Rule: 37.76 ± 22.96 mL MAE.
- Result: A massive improvement, reducing error by ~83%.
Ablation Studies:
- Pose Optimization: Removing pose optimization (fixing views) increased LV volume error from 4.88 mL to 9.32 mL and LA error from 6.40 mL to 14.16 mL, proving that correcting view orientation is critical.
- Ideal vs. Realistic: Even with perturbed (noisy) initial poses, the method performed robustly, approaching the performance of an "ideal" scenario where true poses were known.

5. Significance and Conclusion

This work demonstrates a viable pathway to enhance 2D TTE analysis by leveraging 3D shape priors learned from CT data. By treating the reconstruction as an optimization problem over a neural implicit function, the method can infer the true 3D position of ultrasound planes and reconstruct the full cardiac anatomy, even from sparse and misaligned inputs.

Clinical Impact:

Accuracy: Offers significantly more accurate volumetric quantification than current geometric assumptions (Simpson's rule), particularly for the Left Atrium.
Robustness: The ability to correct for probe misalignment suggests potential for reducing inter-observer variability and improving the reliability of cardiac function assessments.
Future Work: While the method shows promise, the authors note that performance on real 2D TTE images (which contain speckle noise and lower resolution compared to CTA) remains to be validated. Additionally, automated segmentation of 2D TTE images is a prerequisite that requires further development.

In summary, this paper presents a novel, data-driven approach to bridge the gap between 2D imaging limitations and the need for precise 3D cardiac modeling, potentially transforming how cardiac chambers are quantified in routine clinical practice.