Towards Realistic Ultrasound Fetal Brain Imaging Synthesis

TL;DR

This paper explores GAN-based models to generate realistic fetal ultrasound brain images, addressing data scarcity issues for AI training by producing high-resolution synthetic images with promising quality metrics.

Contribution

It introduces diffusion-super-resolution-GAN and transformer-GAN models for fetal ultrasound image synthesis, demonstrating their effectiveness in generating high-quality images.

Findings

GAN models can generate 256x256 fetal ultrasound brain images.

Diffusion-super-resolution-GAN achieved lower FID scores than transformer-GAN.

Stable training losses indicate reliable synthesis of realistic images.

Abstract

Prenatal ultrasound imaging is the first-choice modality to assess fetal health. Medical image datasets for AI and ML methods must be diverse (i.e. diagnoses, diseases, pathologies, scanners, demographics, etc), however there are few public ultrasound fetal imaging datasets due to insufficient amounts of clinical data, patient privacy, rare occurrence of abnormalities in general practice, and limited experts for data collection and validation. To address such data scarcity, we proposed generative adversarial networks (GAN)-based models, diffusion-super-resolution-GAN and transformer-based-GAN, to synthesise images of fetal ultrasound brain planes from one public dataset. We reported that GAN-based methods can generate 256x256 pixel size of fetal ultrasound trans-cerebellum brain image plane with stable training losses, resulting in lower FID values for diffusion-super-resolution-GAN (average 7.04 and lower FID 5.09 at epoch 10) than the FID values of transformer-based-GAN (average 36.02 and lower 28.93 at epoch 60). The results of this work illustrate the potential of GAN-based methods to synthesise realistic high-resolution ultrasound images, leading to future work with other fetal brain planes, anatomies, devices and the need of a pool of experts to evaluate synthesised images. Code, data and other resources to reproduce this work are available at \url{https://github.com/budai4medtech/midl2023}.