Guiding 3D U-nets with signed distance fields for creating 3D models from images

TL;DR

This paper introduces a method that uses signed distance fields to guide 3D U-nets, producing smooth, morphologically accurate 3D models from images, improving over traditional voxelized segmentations.

Contribution

The novel approach integrates signed distance fields into 3D U-net training to generate morphologically consistent models from medical images.

Findings

Produces smoother, more accurate 3D surface reconstructions

Outperforms binary labelmap guidance in surface similarity

Effective on both synthetic and real cardiac data

Abstract

Morphological analysis of the left atrial appendage is an important tool to assess risk of ischemic stroke. Most deep learning approaches for 3D segmentation is guided by binary labelmaps, which results in voxelized segmentations unsuitable for morphological analysis. We propose to use signed distance fields to guide a deep network towards morphologically consistent 3D models. The proposed strategy is evaluated on a synthetic dataset of simple geometries, as well as a set of cardiac computed tomography images containing the left atrial appendage. The proposed method produces smooth surfaces with a closer resemblance to the true surface in terms of segmentation overlap and surface distance.