🧬 biology

Flow-based conditional cardiac anatomy generation for virtual cohorts

This paper introduces CAN-FLOW, a two-step conditional normalizing flow framework that outperforms existing conditional variational autoencoders by generating realistic, metadata-conditioned biventricular cardiac anatomies from UK Biobank data to facilitate the creation of diverse virtual cohorts for clinical research.

Original authors: Konstantinos Kevopoulos, Beatrice Moscoloni, Benjamin Alheit, Cameron Beeche, Julio A. Chirinos, Alexander Heinlein, Mathias Peirlinck

Published 2026-08-11

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Konstantinos Kevopoulos, Beatrice Moscoloni, Benjamin Alheit, Cameron Beeche, Julio A. Chirinos, Alexander Heinlein, Mathias Peirlinck

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are a doctor trying to design a new heart valve. You can't just test it on one person; you need to know how it works on thousands of different people, from tall athletes to shorter seniors, from young adults to the elderly. In the past, scientists tried to build a "digital twin" of a patient's heart by scanning their actual MRI images and turning them into a 3D computer model. But this is like trying to build a library by borrowing a book from every single person in a city: it's slow, expensive, and often impossible because of privacy rules that stop you from sharing those personal scans.

To solve this, researchers are trying to build "virtual cohorts"—groups of fake but realistic hearts generated by computers. Think of it like a video game character creator. You want the computer to generate a heart that looks like it belongs to a 60-year-old male with a specific body type, or a 25-year-old female with a different build. The challenge is making sure these fake hearts aren't just "average" blobs. Real people vary wildly; some hearts are round, some are long, some are huge, and some are tiny. If your computer only makes the "average" heart, your new valve might work on the average person but fail on the ones at the edges of the curve. The goal is to create a generator that understands these differences and can produce a whole crowd of unique, realistic hearts that match specific real-world groups.

This is where the new paper, titled "Flow-based conditional cardiac anatomy generation for virtual cohorts," steps in. The researchers, led by Konstantinos Kevopoulos and Mathias Peirlinck, introduced a new tool called CAN-FLOW. They wanted to see if they could build a computer system that doesn't just make one generic heart, but can create a whole crowd of hearts that change based on who you ask for: "Give me a heart for a 50-year-old male with a BMI of 25," or "Give me one for an 80-year-old female."

To do this, they had to beat the old way of doing things. Before this, most scientists used a method called a conditional Variational Autoencoder (cVAE). You can think of a cVAE like a student trying to memorize a textbook. The student is forced to compress all the information about different hearts into a single, strict "notebook" (a shared mathematical rule) that must look the same for everyone. The problem is that this notebook gets too crowded. To fit everything in, the student has to smooth out the weird, unique details. The result is a group of generated hearts that look okay, but they all end up looking suspiciously similar, like a class of students who all copied the same answer key. They miss the rare but important variations found at the edges of the population.

The authors argue that this old method is too rigid. They propose a new two-step strategy, CAN-FLOW, which is more like a master chef and a sous-chef working together.

Step One (The Chef): First, they teach a computer to look at real heart scans and describe them using a special language called "momenta." This is like translating a complex 3D heart shape into a set of instructions on how to stretch and squish a standard template heart to match the real one. Crucially, at this stage, the computer ignores who the person is (their age, sex, or weight) and just focuses on the shape. It creates a clean, simple map of the heart's geometry.
Step Two (The Sous-Chef): Next, they use a different tool called a conditional normalizing flow. This tool looks at the maps from Step One and learns the rules of how those shapes change based on the person's details. It learns, for example, that as age goes up, the heart shape tends to shift in a specific way, or that males tend to have larger hearts than females.

The team trained this system on 2,208 healthy hearts from the UK Biobank. They then compared their new CAN-FLOW system against the old cVAE methods. The results suggest that CAN-FLOW is much better at capturing the full range of human diversity.

When they looked at the hearts generated by the old cVAE methods, they found the hearts were too uniform. It was as if the computer was only making the "most common" heart and ignoring the outliers. In contrast, CAN-FLOW generated hearts that covered a much wider range of sizes and shapes. The paper shows that CAN-FLOW successfully recreated the "tails" of the distribution—the rare but realistic heart shapes that the old method missed. For instance, when they asked for hearts based on specific ages and body sizes, CAN-FLOW produced groups where the hearts varied naturally, just like in the real world, whereas the old method produced groups where everyone looked almost identical.

The researchers measured this using several tests. They checked if the fake hearts had the right volume and mass, and they found CAN-FLOW matched the real data much more closely. They also looked at the tiny details of the heart's surface, point by point. The old method tended to make hearts that were too smooth and similar to each other, while CAN-FLOW preserved the unique bumps and curves that make every heart distinct.

Importantly, the paper suggests that this approach could be a game-changer for "in silico" (computer-based) clinical trials. Instead of testing a new drug or device on just one or two digital hearts, doctors could test it on a virtual crowd of 600 different hearts, all generated to match a specific demographic. This would help them see if a treatment works for the average person and for the people with unusual heart shapes, making medical testing safer and more inclusive.

However, the authors are careful to note the limits of their work. They only tested this on healthy hearts from a specific group of people in the UK. They haven't yet proven it works on hearts with diseases or on people from different backgrounds. They also only generated static snapshots of the heart (how it looks when it's full of blood), not the moving movie of the heart beating. But as a proof of concept, the paper suggests that by separating the "shape learning" from the "personality learning," we can build virtual crowds that are far more realistic and useful than what we could make before. It's a step toward a future where we can simulate medical treatments on a diverse population of digital humans without needing to scan every single one of them.

Technical Summary: Flow-based Conditional Cardiac Anatomy Generation for Virtual Cohorts

Problem Statement
Cardiac digital twin research is transitioning from subject-specific anatomical replicas to virtual cohorts representing clinically relevant population subgroups. However, the creation of such cohorts is hindered by limited access to representative imaging-derived anatomy datasets due to cohort size constraints, subgroup sparsity, and data-sharing restrictions (privacy, consent, ownership). While synthetic data generation offers a solution, existing conditional generative models, primarily Conditional Variational Autoencoders (cVAEs), face a critical limitation: they couple representation learning with metadata conditioning through a shared, regularized latent prior (typically a standard Gaussian). This shared regularization constrains the variability of latent distributions across distinct metadata groups (e.g., sex, age, body mass index), often leading to the underrepresentation of subgroup-specific anatomical variability, particularly in the tails of the distribution. For virtual cohort construction and in silico clinical trials, preserving these rare but physiologically plausible anatomies is essential.

Methodology: CAN-FLOW
The authors introduce CAN-FLOW (Conditional ANatomy generation with normalizing FLOWs), a two-step framework designed to decouple geometry-only representation learning from metadata-dependent distribution modeling. The framework operates on end-diastolic biventricular surface meshes derived from cardiac magnetic resonance (CMR) images, represented via diffeomorphic momenta using Large-Deformation Diffeomorphic Metric Mapping (LDDMM).

Step 1: Geometry-Only Representation Learning:
- An autoencoder is trained to compress the high-dimensional diffeomorphic momenta ( $\mu_0$ ) into a lower-dimensional latent space ( $z$ ).
- Crucially, this autoencoder is metadata-agnostic. It learns a compact, geometry-only representation without imposing a specific prior distribution or conditioning on subject characteristics (sex, age, BMI) during the encoding/decoding process. This preserves the full anatomical information without forcing it into a shared Gaussian prior.
Step 2: Conditional Distribution Modeling:
- A conditional normalizing flow ( $f_\phi$ ) is trained to model the distribution of the latent representations ( $z$ ) conditioned on metadata ( $c$ = [sex, age, BMI]).
- The flow learns a bijective mapping from the latent space to a base variable $x$ , where the base distribution is a metadata-dependent Gaussian $N(\mu_\phi(c), \Sigma_\phi(c))$ .
- During generation, the flow samples from the learned conditional prior based on specific metadata and maps these samples back to synthetic latent representations, which are then decoded into synthetic momenta and finally transformed back into 3D biventricular surface meshes via geodesic shooting (inverse LDDMM).

Comparison Baselines
The study compares CAN-FLOW against cVAE baselines trained on the same UK Biobank dataset. The cVAEs follow the standard formulation where metadata is provided to both the encoder and decoder, and the latent space is regularized toward a shared Gaussian prior. To ensure a fair comparison, the authors trained cVAEs across a wide range of regularization strengths ( $\beta$ values from $10^{-6}$ to $10^{-1}$ ), reporting the two strongest performing variants ( $\beta = 10^{-2}$ and $\beta = 10^{-3}$ ).

Key Results
The models were evaluated on a healthy subset of 2,208 UK Biobank subjects. CAN-FLOW demonstrated superior performance across five critical requirements for virtual cohort generation:

Anatomical Plausibility: CAN-FLOW generated visually plausible biventricular anatomies that reproduced key structural characteristics (e.g., basal LV shape, chamber dimensions) and exhibited realistic subject-to-subject variation.
Subgroup Variability: In a targeted subgroup (males >58 years), CAN-FLOW reproduced spatial patterns of variation (e.g., in outflow regions and ventricular apex) that closely matched the real cohort. In contrast, cVAE baselines produced overly homogeneous cohorts with uniformly lower spatial variability.
Phenotype Distribution Fidelity: CAN-FLOW achieved the lowest Kullback-Leibler (KL) divergence and competitive Wasserstein distances for clinical phenotypes (LVEDV, RVEDV, myocardial mass, etc.) compared to real distributions. It better reproduced joint phenotype distributions, including the tails and less frequent but physiologically plausible combinations, whereas cVAEs tended to cluster tightly around anatomical means.
Metadata-Dependent Trends: When stratified by sex, age, and BMI, CAN-FLOW better preserved the trends and distributions of clinical phenotypes, particularly capturing the modes and tails of sex-conditioned distributions more accurately than cVAEs.
High-Dimensional Shape Variability: Using Principal Component Analysis (PCA) on shape momenta, CAN-FLOW showed inter-cohort Wasserstein distance ratios closer to 1 (indicating better reproduction of real inter-cohort variability) across leading PCA coefficients. It also achieved approximately 10% higher coverage in point-cloud geometry metrics, indicating a broader span of the real anatomical distribution, while cVAEs favored nearest-neighbor similarity (lower Minimum Matching Distance) at the cost of distributional breadth.

Significance and Claims
The paper claims that CAN-FLOW establishes a shareable framework for generating realistic, stochastically varying, metadata-conditioned biventricular anatomies. Its primary contribution is the separation of representation learning from conditional distribution modeling, which allows the model to preserve the full breadth of anatomical variability (including subgroup-specific heterogeneity and distributional tails) without the constraints imposed by a shared, conditioning-independent prior in cVAEs.

The authors position this framework as a solution to the data-availability barrier in cardiac digital twin research, enabling the construction of virtual cohorts for in silico clinical trials and population-specific model development without the need to share individual image-derived meshes. They note that while the current implementation focuses on static end-diastolic anatomy in healthy populations, the approach offers a pathway toward more accessible and privacy-conscious virtual anatomy cohorts. The study modestly notes that diagnostic applications (e.g., detecting abnormal remodeling) would require further validation in independent diseased cohorts.

Technical Summary: Flow-based Conditional Cardiac Anatomy Generation for Virtual Cohorts

More like this