CardioComposer: Leveraging Differentiable Geometry for Compositional Control of Anatomical Diffusion Models

ICLR 2026 Conference SubmissionAnonymous Authors
Diffusion ModelsComputational GeometryAnatomyDigital TwinsDiffusion Guidance
Abstract:

Generative models of 3D cardiovascular anatomy can synthesize informative structures for clinical research and medical device evaluation, but face a trade-off between geometric controllability and realism. We propose CardioComposer: a programmable, inference-time framework for generating multi-class anatomical label maps based on interpretable ellipsoidal primitives. These primitives represent geometric attributes such as the size, shape, and position of discrete substructures. We specifically develop differentiable measurement functions based on voxel-wise geometric moments, enabling loss-based gradient guidance during diffusion model sampling. We demonstrate that these losses can constrain individual geometric attributes in a disentangled manner and provide compositional control over multiple substructures. Finally, we show that our method is compatible with a wide array of anatomical systems containing non-convex substructures, spanning cardiac, vascular, and skeletal organs.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

CardioComposer introduces a programmable framework for generating multi-class anatomical label maps using interpretable ellipsoidal primitives and differentiable geometric moment functions to guide diffusion sampling. The paper resides in the 'Primitive-Based Compositional Control' leaf, which contains only two papers total (including the original work). This represents a relatively sparse research direction within the broader taxonomy of 13 papers across multiple branches, suggesting the primitive-based approach to anatomical generation is less explored compared to segmentation-mask conditioning methods.

The taxonomy reveals that CardioComposer's approach diverges from neighboring directions in meaningful ways. The sibling leaf 'Topological and Morphological Property Enforcement' focuses on persistent homology and topological features rather than primitive parameterization, while 'Landmark and Skeletal Structure Guidance' uses discrete point sets instead of continuous ellipsoidal representations. The broader 'Segmentation-Guided Anatomical Synthesis' branch (containing six papers across three leaves) represents a more crowded alternative paradigm that conditions on dense masks without explicit geometric parameterization, highlighting CardioComposer's distinct emphasis on interpretable, modular geometric handles.

Among 29 candidates examined, the contribution-level analysis shows varied novelty profiles. The differentiable geometric measurement functions (10 candidates examined, 0 refutable) and inference-time guidance framework (9 candidates examined, 0 refutable) appear to have limited direct prior work within the search scope. However, the compositional control framework for multi-part anatomical constraints (10 candidates examined, 1 refutable) shows at least one overlapping candidate, suggesting some existing work addresses multi-structure compositional generation. The limited search scope means these findings reflect top-K semantic matches rather than exhaustive coverage.

Given the sparse primitive-based leaf and the limited 29-candidate search, CardioComposer appears to occupy a relatively underexplored niche within anatomical diffusion modeling. The single sibling paper and the refutation of one contribution among 29 candidates suggest moderate novelty, though the analysis cannot rule out relevant work outside the top-K semantic neighborhood. The framework's extension to cardiac, vascular, and skeletal systems may represent incremental breadth rather than fundamental methodological departure from the sibling work.

Taxonomy

Core-task Taxonomy Papers
13
3
Claimed Contributions
29
Contribution Candidate Papers Compared
1
Refutable Paper

Research Landscape Overview

Core task: Compositional geometric control of anatomical diffusion models. The field centers on generating anatomically plausible medical images through diffusion models that respect geometric and structural constraints. The taxonomy reveals several complementary directions: geometric constraint mechanisms that enforce spatial and shape priors during generation, segmentation-guided approaches that leverage anatomical masks to steer synthesis, multimodal conditioning architectures that integrate diverse input modalities (text, sketches, landmarks), semantic augmentation methods that produce domain-agnostic or label-preserving synthetic data, and specialized diffeomorphic mapping techniques for neuroanatomical registration. Representative works span cardiac imaging (Heartbeat[3], Coronary Anatomy Diffusion[6]), broader anatomical synthesis (Anatomica[5]), and surgical or polyp detection scenarios (Surgical Scene Augmentation[1], Polyp Detection Diffusion[7]), illustrating how different branches address organ-specific versus general anatomical generation challenges. A particularly active line of work explores primitive-based compositional control, where generation is decomposed into interpretable geometric elements such as landmarks, contours, or parametric shape descriptors. CardioComposer[0] exemplifies this approach by enabling fine-grained compositional manipulation of cardiac structures through geometric primitives, closely related to CardioComposer Flexible[8] which extends similar compositional strategies. This contrasts with segmentation-guided methods like Anatomically-Controllable Generation[11] that rely on dense mask conditioning, and with semantic augmentation frameworks such as Semantic Data Augmentation[4] that prioritize label consistency over explicit geometric control. Meanwhile, diffeomorphic techniques like Diffeomorphic Neuroanatomy[2] focus on smooth, topology-preserving transformations rather than direct synthesis. The original paper sits within the primitive-based cluster, emphasizing interpretable geometric handles for compositional control, distinguishing itself from mask-driven approaches by offering more modular, editable generation pathways that align with clinical workflows requiring precise anatomical adjustments.

Claimed Contributions

Differentiable geometric measurement functions for anatomical characterization

The authors develop differentiable functions that measure voxel-wise geometric moments to characterize anatomical substructures. These functions compute size via zeroth-order moments, position via first-order moments, and shape via scale-normalized second-order moments, enabling gradient-based optimization during diffusion sampling.

10 retrieved papers
Inference-time guidance framework for controlling substructure geometry

The authors present an inference-time method that uses gradients from geometric loss functions to guide unconditional diffusion models. This approach enables independent or joint control of substructure attributes without retraining the model, where substructures can consist of one tissue class or unions of multiple classes.

9 retrieved papers
Compositional control framework for multi-part anatomical constraints

The authors demonstrate that their framework supports compositional generation by combining multiple substructure-specific geometric losses. This enables complex anatomical constraints across arbitrary numbers of substructures, including non-convex geometries with branching or curved features.

10 retrieved papers
Can Refute

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

Differentiable geometric measurement functions for anatomical characterization

The authors develop differentiable functions that measure voxel-wise geometric moments to characterize anatomical substructures. These functions compute size via zeroth-order moments, position via first-order moments, and shape via scale-normalized second-order moments, enabling gradient-based optimization during diffusion sampling.

Contribution

Inference-time guidance framework for controlling substructure geometry

The authors present an inference-time method that uses gradients from geometric loss functions to guide unconditional diffusion models. This approach enables independent or joint control of substructure attributes without retraining the model, where substructures can consist of one tissue class or unions of multiple classes.

Contribution

Compositional control framework for multi-part anatomical constraints

The authors demonstrate that their framework supports compositional generation by combining multiple substructure-specific geometric losses. This enables complex anatomical constraints across arbitrary numbers of substructures, including non-convex geometries with branching or curved features.