Binary segmentation of medical images using implicit spline representations and deep learning

Oliver J.D Barrowclough,Georg Muntingh,Varatharajan Nainamalai,Ivar Stangeby

doi:10.1016/j.cagd.2021.101972

Oliver J.D Barrowclough, Georg Muntingh + Show 2 more

Open Access

https://doi.org/10.1016/j.cagd.2021.101972

Copy DOI

Abstract

We propose a novel approach to image segmentation based on combining implicit spline representations with deep convolutional neural networks. This is done by predicting the control points of a bivariate spline function whose zero-set represents the segmentation boundary. We adapt several existing neural network architectures and design novel loss functions that are tailored towards providing implicit spline curve approximations. The method is evaluated on a congenital heart disease computed tomography medical imaging dataset. Experiments are carried out by measuring performance in various standard metrics for different networks and loss functions. We determine that splines of bidegree (1,1) with 128×128 coefficient resolution performed optimally for 512×512 resolution CT images. For our best network, we achieve an average volumetric test Dice score of close to 92%, which reaches the state of the art for this congenital heart disease dataset.

Highlights

Image segmentation is the process of partitioning an image of pixels into different regions according to shared attributes
We focus on the problem of segmenting computed tomography (CT) images from a dataset of congenital heart disease (CHD) patients Xu et al (2019b)
The method is applicable to general image segmentation problems, and we show that it reaches state-of-the-art results on a CHD CT image dataset

Summary

Introduction

Image segmentation is the process of partitioning an image of pixels into different regions according to shared attributes. There have been some attempts to obtain ML-based automatic CT and MRI cardiac image segmentation for full blood volume or for heart chambers. Xu et al (2019b) used a UNet-based deep learning network for CHD CT images of 68 volumes and obtained an average Dice score of 0.7843 and 0.773 for blood volume and myocardium, respectively (2D UNet for blood volume segmentation, 3D UNet for chambers and myocardium segmentation). Payer et al (2018) used a 3D UNet architecture model with bounding box around all heart structures, and obtained a mean Dice score of 0.889 for the whole heart CT image segmentation. Habijan et al (2019) used a 3D UNet architecture CNN model with principal component analysis as a data augmentation technique, and obtained an average Dice score of 0.89 for the whole heart CT image segmentation

Methods

Results

Conclusion