CardioComposer: Flexible and Compositional Anatomical Structure Generation with Disentangled Geometric Guidance

Karim Kadry, Shoaib Goraya, Ajay Manicka, Abdalla Abdelwahed, Farhad R. Nezami, Elazer R. Edelman · 2025

Read CardioComposer: Flexible and Compositional Anatomical Structure Generation with Disentangled Geometric Guidance by Karim Kadry, Shoaib Goraya, Ajay Manicka, Abdalla Abdelwahed, Farhad R. Nezami, Elazer R. Edelman (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-08015.txt.

Introduction

Synthetic 3D heart and vessel anatomy with real control over shape is offered by Karim Kadry and 6 co-authors in a 2025 medical imaging paper. Generative models of cardiovascular anatomy help clinical research and device testing, but they usually force a trade: either realistic shapes that cannot be adjusted, or adjustable shapes that look artificial.

Their framework, CardioComposer, works at inference time and builds multi-class anatomical label maps from simple ellipsoidal building blocks. Those blocks encode size, shape, and position of anatomical parts in readable terms. Differentiable measurements based on geometric moments let gradients steer a diffusion model's sampling toward the requested anatomy. Programmable anatomy could supply test cases for devices and training data for rare conditions.

Catalog record

CREATORKarim Kadry, Shoaib Goraya, Ajay Manicka, Abdalla Abdelwahed, Farhad R. Nezami, Elazer R. Edelman
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE48 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-08015

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

CardioComposer: Flexible and Compositional Anatomical Structure Generation with Disentangled Geometric Guidance Karim Kadry, Shoaib Goraya, Ajay Manicka, Abdalla Abdelwahed, Farhad R. Nezami, Elazer R. Edelman Abstract Generative models of 3D anatomy, when integrated with biophysical simulators, enable the study of structure-function relationships for clinical research and medical device design. However, current models face a trade-off between controllability and anatomical realism. We propose a programmable and compositional framework for guiding unconditional diffusion models of human anatomy using interpretable ellipsoidal primitives embedded in 3D space. Our method involves the selection of certain tissues within multi-tissue segmentation maps, upon which we apply geometric moment losses to guide the reverse diffusion process. This framework supports the independent control over size, shape, and position, as well as the composition of multi-component constraints during inference. 1 Introduction Three-dimensional segmentations of human anatomy power a variety of physics-based simulation platforms. For example, virtual cohorts of anatomy can be used for virtual clinical trials to evaluate and optimize novel medical devices and imaging systems [ 1 , 2 , 3 ] . Additionally, biophysical simulations can generate insights in the context of both computational physiology research [ 4 , 5 , 6 ] , and surgical training [ 7 ] . Anatomical segmentations can also be used to augment machine-learning workflows through simulating the formation of synthetic images, either through physics-based models [ 8 , 9 ] , domain randomization [ 10 , 11 ] , or using anatomy-controlled generative models [ 12 , 13 ] . Figure 1: We present a guidance framework to constrain diffusion models of multi-label anatomical segmentations based on simple geometric features . We specifically guide the model with interpretable and editable features that can represent ellipsoids in 3D space, such as size, shape, and position ( panel A ). Our approach does not require model re-training, enables the independent generation of anatomy based on various geometric features ( panels B-D ), and supports multi-component compositional generation ( panels E-G ). For panels B-G, gray and blue voxels represent components that are unconstrained and constrained, respectively, and purple ellipsoids indicate a strong overlap between target and sample ellipsoids . Using generative models of anatomy as a surrogate data source for simulation platforms offers several advantages over using patient-data alone. For example, conditional generation can augment computational trial cohorts with anatomical variants that are both novel and rare [ 14 ] . Moreover, generative editing methods such as inpainting can precisely modify existing patient geometry to create anatomically plausible variations [ 15 , 16 ] . These ”digital siblings” can be used with biophysical simulators to model counterfactual scenarios that elucidate the relationship between anatomical form and function. However, unlike generative modeling of 3D shapes for artistic purposes, generating anatomical models for biophysical simulations presents several unique challenges. The first challenge relates to scale-critical features, where minor geometric variations on the order of mere millimeters can cause major fluctuations in physiological behavior [ 17 , 18 , 19 ] . Second, anatomical structure exerts attribute-specific effects, where features such as size and position play differing roles in determining biophysical outcomes [ 20 ] . Lastly, multiple components or tissues interact in a compositional manner [ 20 , 21 , 22 ] , where the relative arrangement and properties of adjacent structures collectively influence simulation results. These challenges highlight the essential functionalities for anatomical generative models, namely: that they can precisely control critical geometric attributes, independently manipulate different attributes such as size or shape, and control the geometry of multiple components in a compositional manner. To address these design requirements, we present CardioComposer, a novel framework for guiding unconditional diffusion models using interpretable ellipsoidal primitives that represent geometric attributes such as size, shape, and position ( Figure 1 A). Our guidance framework allows for independent control of size, shape, and position ( Figure 1 B-D), as well as the ability to compose geometric constraints for arbitrary numbers of anatomical components ( Figure 1 E-G). Our key insight is that a linear combination of simple loss functions applied to target and predicted geometric features can jointly constrain anatomical generation. We demonstrate this method on multi-tissue cardiovascular segmentations, which exhibit a wide array of structures such as star-shaped chambers, and tubular vasculature. Our key contributions include the following: • We reformulate geometrically constrained synthesis of 3D anatomical segmentations in terms of guiding unconditional diffusion models with a differentiable geometric loss at inference. • We develop a composite geometric loss function enabling independent control over size, shape, and position. • We apply such geometric losses to individual tissues within the anatomical segmentation, enabling multi-part compositional generation. Figure 2: Our method involves applying a geometric guidance correction step for every denoising iteration . Left : the noised latent \mathbf{z}_{\sigma} is passed through the diffusion model and VAE decoder to produce a clean voxel space prediction \mathbf{\hat{x}}_{0} ( Section 4.2 ). Middle : The segmentation is filtered for components of interest \mathbf{\Omega} , and geometric moments \mathcal{G} are extracted per component ( Section 4.3 ). Right : the measured geometric moments \mathcal{G} are compared to target moments \bar{\mathcal{G}} through geometric moment losses, and are aggregated to produce a final loss that can be used to guide sampling by taking the gradient ( Section 4.4 ). Bottom The gradient derived from the geometric loss is then used to adjust the denoising step. 2 Background Traditional Morphometric Modeling for Anatomy. Anatomical modeling involves digitally quantifying and representing anatomical structures through geometric measurements. Traditional morphometric approaches characterize anatomy through intrinsic factors such as length, area, volume, and shape, as well as extrinsic factors such as position and orientation. These measurements are often performed in specialized coordinate systems adapted to specific anatomical contexts; curvilinear coordinates for vascular trees [ 23 ] , polar coordinates for valves [ 24 ] , and custom coordinate systems for biventricular geometries [ 25 ] . Geometric measurements enable various clinical applications, including cardiovascular risk stratification [ 26 , 27 ] , orthopedic diagnosis [ 28 ] , and radiomics-based prognosis [ 29 ] . However, traditional morphometric modeling faces two key challenges: it cannot capture complex nonlinear correlations between features, and multiple distinct anatomies may map to the same high-level measurements. To address these limitations, we propose a framework in which an unconditional diffusion model is controlled by traditional morphometric features (size, shape, position) to generate a wide variety of realistic anatomical structures, providing an approachable anatomical modeling interface for both clinical and engineering workflows that involve numerical simulation. Generative Models for Numerical Simulation. Anatomical modeling serves as a crucial tool for both studying form-function relationships and enabling physical simulations for scientific discovery and medical device design. Current approaches span a spectrum of complexity and control, yet each faces fundamental limitations. Simple geometric primitives, such as cylinders for coronary arteries [ 30 , 31 , 32 ] and truncated ellipsoids for cardiac chambers [ 33 ] offer parametric control but fail to capture anatomical realism. Data-driven approaches like statistical shape models and autoencoders [ 34 , 35 , 36 , 37 , 14 , 38 ] can represent anatomy in terms of a global shape vector, where each element corresponds to composite shape factors. Data-driven approaches have recently

Read the complete text