Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

NUDF: Neural Unsigned Distance Fields for high resolution 3D medical image segmentation

Kristine Sørensen, Oscar Camara, Ole de Backer, Klaus Kofoed, Rasmus Paulsen

TL;DR

This work introduces Neural Unsigned Distance Fields (NUDF) to perform high-resolution 3D medical image segmentation by learning a continuous distance function directly from CT images, enabling open-surface mesh reconstructions. The method uses a CNN encoder to extract image features and a pointwise FCNN decoder to predict the unsigned distance at queried points, guided by a Shape Diameter–based sampling strategy. NUDF demonstrates superior surface fidelity over a traditional 3D U-net on Left Atrial Appendage segmentation, achieving mesh-quality metrics close to voxel spacing while handling complex, open geometries with lower memory demands. The approach offers a practical path to high-resolution, visualization- and simulation-ready meshes in clinical imaging, with potential applicability to other anatomies requiring detailed surface models.

Abstract

Medical image segmentation is often considered as the task of labelling each pixel or voxel as being inside or outside a given anatomy. Processing the images at their original size and resolution often result in insuperable memory requirements, but downsampling the images leads to a loss of important details. Instead of aiming to represent a smooth and continuous surface in a binary voxel-grid, we propose to learn a Neural Unsigned Distance Field (NUDF) directly from the image. The small memory requirements of NUDF allow for high resolution processing, while the continuous nature of the distance field allows us to create high resolution 3D mesh models of shapes of any topology (i.e. open surfaces). We evaluate our method on the task of left atrial appendage (LAA) segmentation from Computed Tomography (CT) images. The LAA is a complex and highly variable shape, being thus difficult to represent with traditional segmentation methods using discrete labelmaps. With our proposed method, we are able to predict 3D mesh models that capture the details of the LAA and achieve accuracy in the order of the voxel spacing in the CT images.

NUDF: Neural Unsigned Distance Fields for high resolution 3D medical image segmentation

TL;DR

This work introduces Neural Unsigned Distance Fields (NUDF) to perform high-resolution 3D medical image segmentation by learning a continuous distance function directly from CT images, enabling open-surface mesh reconstructions. The method uses a CNN encoder to extract image features and a pointwise FCNN decoder to predict the unsigned distance at queried points, guided by a Shape Diameter–based sampling strategy. NUDF demonstrates superior surface fidelity over a traditional 3D U-net on Left Atrial Appendage segmentation, achieving mesh-quality metrics close to voxel spacing while handling complex, open geometries with lower memory demands. The approach offers a practical path to high-resolution, visualization- and simulation-ready meshes in clinical imaging, with potential applicability to other anatomies requiring detailed surface models.

Abstract

Medical image segmentation is often considered as the task of labelling each pixel or voxel as being inside or outside a given anatomy. Processing the images at their original size and resolution often result in insuperable memory requirements, but downsampling the images leads to a loss of important details. Instead of aiming to represent a smooth and continuous surface in a binary voxel-grid, we propose to learn a Neural Unsigned Distance Field (NUDF) directly from the image. The small memory requirements of NUDF allow for high resolution processing, while the continuous nature of the distance field allows us to create high resolution 3D mesh models of shapes of any topology (i.e. open surfaces). We evaluate our method on the task of left atrial appendage (LAA) segmentation from Computed Tomography (CT) images. The LAA is a complex and highly variable shape, being thus difficult to represent with traditional segmentation methods using discrete labelmaps. With our proposed method, we are able to predict 3D mesh models that capture the details of the LAA and achieve accuracy in the order of the voxel spacing in the CT images.

Paper Structure

This paper contains 11 sections, 3 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Data
  3. Methods
  4. Neural Implicit Function learning
  5. Sampling strategy
  6. Meshing unsigned distance fields
  7. Implementation details
  8. Experiments
  9. Discussion and conclusion
  10. Compliance with ethical standards
  11. Acknowledgements

Figures (3)

  • Figure 1: Overview of the proposed method for creating 3D mesh models from high resolution input images. The input image is processed through a series of 3D convolutions and maxpoolings to produce feature maps $G_{1..N}$. The feature maps are sampled at continuous point coordinates (red points) and inputted to a fully connected neural network (FCNN) predicting the distance from the point to the surface. By repeating this for many points we can learn a distance field representing the anatomy of interest, from which a dense point cloud and/or a triangulated mesh can be extracted.
  • Figure 2: Left) A 2D illustration of the shape diameter (SD) calculation, where a cone of rays (grey area) is cast from each point in the direction of the normal (red arrow). Right) Example of the SD evaluated with the original and with flipped normals.
  • Figure 3: Examples of meshes from manual segmentation, NUDF and a standard 3D U-net. The examples are ordered from left to right as best to worst chamfer distance on our proposed method.