Generating Synthetic Labeled Data From Existing Anatomical Models: An Example With Echocardiography Segmentation.

Andrew Gilbert,Pablo Lamata,Eigil Samset,Cristobal Rodero,Kristin Mcleod,Maciej Marciniak

doi:10.1109/tmi.2021.3051806

Andrew Gilbert, Pablo Lamata + Show 4 more

Open Access

https://doi.org/10.1109/tmi.2021.3051806

Copy DOI

Abstract

Deep learning can bring time savings and increased reproducibility to medical image analysis. However, acquiring training data is challenging due to the time-intensive nature of labeling and high inter-observer variability in annotations. Rather than labeling images, in this work we propose an alternative pipeline where images are generated from existing high-quality annotations using generative adversarial networks (GANs). Annotations are derived automatically from previously built anatomical models and are transformed into realistic synthetic ultrasound images with paired labels using a CycleGAN. We demonstrate the pipeline by generating synthetic 2D echocardiography images to compare with existing deep learning ultrasound segmentation datasets. A convolutional neural network is trained to segment the left ventricle and left atrium using only synthetic images. Networks trained with synthetic images were extensively tested on four different unseen datasets of real images with median Dice scores of 91, 90, 88, and 87 for left ventricle segmentation. These results match or are better than inter-observer results measured on real ultrasound datasets and are comparable to a network trained on a separate set of real images. Results demonstrate the images produced can effectively be used in place of real data for training. The proposed pipeline opens the door for automatic generation of training data for many tasks in medical imaging as the same process can be applied to other segmentation or landmark detection tasks in any modality. The source code and anatomical models are available to other researchers.11https://adgilbert.github.io/data-generation/

Highlights

M EDICAL imaging provides a window to capture the structure and function of internal anatomies
2) We demonstrate the utility of the pipeline by building annotated synthetic 2D echo images from cardiac models. We show these synthetic images can be used for training deep learning algorithms, by demonstrating accurate segmentation without any real labeled images
While the segmentation network can be included within the CycleGAN for end-toend training [18], [26], we found the segmentation network was able to consistently achieve very good results on the synthetic images in preliminary results and did not find value in including this as a loss term within the transformation process

Summary

Introduction

M EDICAL imaging provides a window to capture the structure and function of internal anatomies. Imaging modalities such as ultrasound, computed tomography (CT) or magnetic resonance imaging (MRI) can be used to measure physical and physiological parameters. Accurate automation of these measurements would provide significant time-savings for clinical practitioners. In cardiovascular ultrasound (echocardiography or ‘echo’), inter-observer errors for labeling common measurements can range from 4-22% even for experienced cardiologists [1], [2]. A second problem when building datasets to automate tasks in medical imaging is labeling is time-consuming and expensive since quality annotations require experienced medical professionals. Manual labels are inflexible and adapting them based on new insights requires a significant amount of time

Objectives

Methods

Results

Discussion

Conclusion