Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Steerable Anatomical Shape Synthesis with Implicit Neural Representations

Bram de Wilde, Max T. Rietberg, Guillaume Lajoinie, Jelmer M. Wolterink

TL;DR

The paper tackles the challenge of generating anatomically accurate, steerable 3D shapes for virtual imaging trials by leveraging implicit neural representations (INRs) to model signed distance fields conditioned on a latent code. By using the thyroid as a case study, it demonstrates how conditioning on volume, isthmus area, and symmetry enables targeted shape editing and population-level synthesis, even across topological variations. A correlation loss is proposed to improve disentanglement between fixed anatomical features and trainable latent factors, enhancing steerability. The approach supports generating patient cohorts with specific anatomical traits and permits editing existing thyroid shapes with largely independent control over features, offering a practical tool for medical imaging studies and method validation.

Abstract

Generative modeling of anatomical structures plays a crucial role in virtual imaging trials, which allow researchers to perform studies without the costs and constraints inherent to in vivo and phantom studies. For clinical relevance, generative models should allow targeted control to simulate specific patient populations rather than relying on purely random sampling. In this work, we propose a steerable generative model based on implicit neural representations. Implicit neural representations naturally support topology changes, making them well-suited for anatomical structures with varying topology, such as the thyroid. Our model learns a disentangled latent representation, enabling fine-grained control over shape variations. Evaluation includes reconstruction accuracy and anatomical plausibility. Our results demonstrate that the proposed model achieves high-quality shape generation while enabling targeted anatomical modifications.

Steerable Anatomical Shape Synthesis with Implicit Neural Representations

TL;DR

The paper tackles the challenge of generating anatomically accurate, steerable 3D shapes for virtual imaging trials by leveraging implicit neural representations (INRs) to model signed distance fields conditioned on a latent code. By using the thyroid as a case study, it demonstrates how conditioning on volume, isthmus area, and symmetry enables targeted shape editing and population-level synthesis, even across topological variations. A correlation loss is proposed to improve disentanglement between fixed anatomical features and trainable latent factors, enhancing steerability. The approach supports generating patient cohorts with specific anatomical traits and permits editing existing thyroid shapes with largely independent control over features, offering a practical tool for medical imaging studies and method validation.

Abstract

Generative modeling of anatomical structures plays a crucial role in virtual imaging trials, which allow researchers to perform studies without the costs and constraints inherent to in vivo and phantom studies. For clinical relevance, generative models should allow targeted control to simulate specific patient populations rather than relying on purely random sampling. In this work, we propose a steerable generative model based on implicit neural representations. Implicit neural representations naturally support topology changes, making them well-suited for anatomical structures with varying topology, such as the thyroid. Our model learns a disentangled latent representation, enabling fine-grained control over shape variations. Evaluation includes reconstruction accuracy and anatomical plausibility. Our results demonstrate that the proposed model achieves high-quality shape generation while enabling targeted anatomical modifications.

Paper Structure

This paper contains 15 sections, 4 equations, 4 figures.

Table of Contents

  1. Introduction
  2. Methods
  3. Model
  4. Training
  5. Disentanglement
  6. Fixed conditioning
  7. Correlation loss
  8. Inference
  9. Data
  10. Validation
  11. Results
  12. Reconstruction quality
  13. Generation quality
  14. Steerability
  15. Discussion

Figures (4)

  • Figure 1: Examples from the dataset illustrating the variety in thyroid anatomy. From left to right: (1) a typical connected thyroid, (2) a typical split thyroid, (3) a very large thyroid, (4) a very small thyroid, (5) a highly asymmetric thyroid.
  • Figure 2: Comparison between volume, isthmus area and symmetry for the training data (blue) and 1000 randomly generated meshes using the baseline model (red).
  • Figure 3: Correlation between conditioned and generated anatomical features for the fixed model (top row) and the correlation model (bottom row). The inset shows the Pearson correlation coefficient.
  • Figure 4: Editing a training mesh (middle column, green) by independently varying volume (red), isthmus area (blue) and symmetry (grey). The plots show each anatomical feature for each of the rows, demonstrating that features can be independently varied.