Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fast Medical Shape Reconstruction via Meta-learned Implicit Neural Representations

Gaia Romana De Paolis, Dimitrios Lenis, Johannes Novotny, Maria Wimmer, Astrid Berg, Theresa Neubauer, Philip Matthias Winter, David Major, Ariharasudhan Muthusami, Gerald Schröcker, Martin Mienkina, Katja Bühler

TL;DR

This work tackles fast 3D medical shape reconstruction from sparse segmentations by learning a shared INR initialization via meta-learning. By applying MAML to shape priors, the method enables rapid adaptation to unseen anatomies with a single optimization step, achieving accuracy comparable to state-of-the-art baselines while reducing inference time to around $0.1$ s. The approach generalizes across input configurations (sparse slices, different orientations) and demonstrates transferability to shape domains not seen during training, as shown on vertebra, pancreas, and heart datasets. The resulting framework has clear potential for real-time clinical tasks such as surgical navigation and interactive planning, with room for further enhancements in architecture and meta-learning strategy.

Abstract

Efficient and fast reconstruction of anatomical structures plays a crucial role in clinical practice. Minimizing retrieval and processing times not only potentially enhances swift response and decision-making in critical scenarios but also supports interactive surgical planning and navigation. Recent methods attempt to solve the medical shape reconstruction problem by utilizing implicit neural functions. However, their performance suffers in terms of generalization and computation time, a critical metric for real-time applications. To address these challenges, we propose to leverage meta-learning to improve the network parameters initialization, reducing inference time by an order of magnitude while maintaining high accuracy. We evaluate our approach on three public datasets covering different anatomical shapes and modalities, namely CT and MRI. Our experimental results show that our model can handle various input configurations, such as sparse slices with different orientations and spacings. Additionally, we demonstrate that our method exhibits strong transferable capabilities in generalizing to shape domains unobserved at training time.

Fast Medical Shape Reconstruction via Meta-learned Implicit Neural Representations

TL;DR

This work tackles fast 3D medical shape reconstruction from sparse segmentations by learning a shared INR initialization via meta-learning. By applying MAML to shape priors, the method enables rapid adaptation to unseen anatomies with a single optimization step, achieving accuracy comparable to state-of-the-art baselines while reducing inference time to around s. The approach generalizes across input configurations (sparse slices, different orientations) and demonstrates transferability to shape domains not seen during training, as shown on vertebra, pancreas, and heart datasets. The resulting framework has clear potential for real-time clinical tasks such as surgical navigation and interactive planning, with room for further enhancements in architecture and meta-learning strategy.

Abstract

Efficient and fast reconstruction of anatomical structures plays a crucial role in clinical practice. Minimizing retrieval and processing times not only potentially enhances swift response and decision-making in critical scenarios but also supports interactive surgical planning and navigation. Recent methods attempt to solve the medical shape reconstruction problem by utilizing implicit neural functions. However, their performance suffers in terms of generalization and computation time, a critical metric for real-time applications. To address these challenges, we propose to leverage meta-learning to improve the network parameters initialization, reducing inference time by an order of magnitude while maintaining high accuracy. We evaluate our approach on three public datasets covering different anatomical shapes and modalities, namely CT and MRI. Our experimental results show that our model can handle various input configurations, such as sparse slices with different orientations and spacings. Additionally, we demonstrate that our method exhibits strong transferable capabilities in generalizing to shape domains unobserved at training time.
Paper Structure (19 sections, 3 equations, 4 figures, 7 tables)

This paper contains 19 sections, 3 equations, 4 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Problem Formulation
  4. Fast Shape Reconstruction via Meta-Learning
  5. Shape Prior Meta-Learning
  6. Shape Reconstruction
  7. Experimental Setup
  8. Datasets
  9. Baselines
  10. Implementation Details
  11. Evaluation Setup
  12. Results
  13. Fast Convergence
  14. Reconstruction from Sagittal Slices
  15. Generalizing from different reconstruction tasks
  16. ...and 4 more sections

Figures (4)

  • Figure 1: Our model based on meta-learned INRs can reconstruct anatomical shapes in a single optimization step, achieving accuracy comparable to both baselines (B. (i), B. (ii)), while also being significantly faster at inference time.
  • Figure 2: Average $DSC$ over reconstruction time measured at step 1, 50, 100 and at convergence. The shaded areas represent the corresponding standard deviations. The meta-learned initialization (ours) allows to fast ($\sim 0.1s$) recover a new sample, with accuracy on the level of B.(i) and B.(ii) at convergence.
  • Figure 3: Qualitative comparisons of reconstruction performance at different optimization steps and at convergence.
  • Figure 4: Transfer learning qualitative results example: reconstruction of heart and pancreas shapes from model trained on a different domain shape (vertebra).