Generative Adversarial Networks to Improve Fetal Brain Fine-Grained Plane Classification

Abstract

Generative adversarial networks (GANs) have been recently applied to medical imaging on different modalities (MRI, CT, X-ray, etc). However there are not many applications on ultrasound modality as a data augmentation technique applied to downstream classification tasks. This study aims to explore and evaluate the generation of synthetic ultrasound fetal brain images via GANs and apply them to improve fetal brain ultrasound plane classification. State of the art GANs stylegan2-ada were applied to fetal brain image generation and GAN-based data augmentation classifiers were compared with baseline classifiers. Our experimental results show that using data generated by both GANs and classical augmentation strategies allows for increasing the accuracy and area under the curve score.

Keywords: deep learning; generative adversarial networks; ultrasound image classification.

Figure 1

Fetal brain plane images used in this study [8].

Figure 2

Generation of Trans-thalamic images for some random seeds and different $$\psi$$. Same 25 seeds were applied to each grid giving the same 25 brain plane generation for three $$\psi$$ values and no truncation. (a) $$\psi =0.3$$; (b) $$\psi =0.5$$; (c) $$\psi =0.7$$; (d) $$\psi =1$$ (no truncation).

Figure 3

Generation of Trans-ventricular images for some random seeds and different $$\psi$$. Same 25 seeds were applied to each grid giving the same 25 brain plane generation for three $$\psi$$ values and no truncation. (a) $$\psi =0.3$$; (b) $$\psi =0.5$$; (c) $$\psi =0.7$$; (d) $$\psi =1$$ (no truncation).

Figure 3

Generation of Trans-ventricular images for some random seeds and different $$\psi$$. Same 25 seeds were applied to each grid giving the same 25 brain plane generation for three $$\psi$$ values and no truncation. (a) $$\psi =0.3$$; (b) $$\psi =0.5$$; (c) $$\psi =0.7$$; (d) $$\psi =1$$ (no truncation).

Figure 4

Accuracy (blue, with max and min) and AUC (green, with max and min) for experiments with $$\psi =0.3$$, $$\psi =0.5$$, $$\psi =0.7$$ and $$\psi =1$$ (no truncation). Horizontal lines represent the baseline accuracy and AUC (without GAN data augmentation).

Figure 5

Accuracy (blue, with max and min) and AUC (green, with max and min) for replacement experiments for no truncation. Horizontal lines represent the baseline accuracy and AUC (without GAN data augmentation).