Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fitting Skeletal Models via Graph-based Learning

Nicolás Gaggion, Enzo Ferrante, Beatriz Paniagua, Jared Vicory

TL;DR

This work introduces a graph-based skeletal learning pipeline (HybridVNet) to predict fixed-topology s-reps from dense segmentation masks, addressing the slow, parameter-heavy fitting of template-based s-reps. The method encodes the object interior as a skeleton–boundary graph G=(V,A,X) via a variational autoencoder with a graph-convolutional decoder, optimized with reconstruction and KL divergence losses. Evaluations on synthetic ellipsoids and real hippocampus data demonstrate accurate skeletal representations with fast inference times, achieving comparable fidelity to deformable template fitting while significantly reducing computation time. The approach offers a scalable, robust avenue for medical shape analysis, with potential gains by incorporating explicit medialness and boundary-orthogonality terms into the loss.

Abstract

Skeletonization is a popular shape analysis technique that models an object's interior as opposed to just its boundary. Fitting template-based skeletal models is a time-consuming process requiring much manual parameter tuning. Recently, machine learning-based methods have shown promise for generating s-reps from object boundaries. In this work, we propose a new skeletonization method which leverages graph convolutional networks to produce skeletal representations (s-reps) from dense segmentation masks. The method is evaluated on both synthetic data and real hippocampus segmentations, achieving promising results and fast inference.

Fitting Skeletal Models via Graph-based Learning

TL;DR

This work introduces a graph-based skeletal learning pipeline (HybridVNet) to predict fixed-topology s-reps from dense segmentation masks, addressing the slow, parameter-heavy fitting of template-based s-reps. The method encodes the object interior as a skeleton–boundary graph G=(V,A,X) via a variational autoencoder with a graph-convolutional decoder, optimized with reconstruction and KL divergence losses. Evaluations on synthetic ellipsoids and real hippocampus data demonstrate accurate skeletal representations with fast inference times, achieving comparable fidelity to deformable template fitting while significantly reducing computation time. The approach offers a scalable, robust avenue for medical shape analysis, with potential gains by incorporating explicit medialness and boundary-orthogonality terms into the loss.

Abstract

Skeletonization is a popular shape analysis technique that models an object's interior as opposed to just its boundary. Fitting template-based skeletal models is a time-consuming process requiring much manual parameter tuning. Recently, machine learning-based methods have shown promise for generating s-reps from object boundaries. In this work, we propose a new skeletonization method which leverages graph convolutional networks to produce skeletal representations (s-reps) from dense segmentation masks. The method is evaluated on both synthetic data and real hippocampus segmentations, achieving promising results and fast inference.
Paper Structure (13 sections, 2 figures, 2 tables)

This paper contains 13 sections, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Skeletal Representations
  4. S-reps via HybridVNet
  5. Results
  6. Experiment Setup
  7. Data Collection and Partitioning
  8. Model Training
  9. Evaluation Metrics
  10. Synthetic experiment
  11. Hippocampus experiment
  12. Conclusion
  13. Compliance with Ethical Standards

Figures (2)

  • Figure 1: (Left) Medial axis for a 2D shape, (Right) s-rep for a hippocampus surface with yellow lines as the spoke vectors.
  • Figure 2: Model architecture: The presented model utilizes a variational encoder-decoder architecture to create a graph representation of an s-rep derived from a binary input image. The encoder comprises a 3D convolutional neural network, producing $\boldsymbol{\mu}$ and $\boldsymbol{\sigma}$ vectors which are sampled, yielding a latent representation denoted as $\mathbf{z}$. This latent code is subsequently goes through a fully connected layer and is reshaped to establish the primary node attributes for the graph convolutional decoder. Leveraging these initial node attributes, the decoder generates the conclusive graph representation of the s-rep.