WGAN-GP based Framework for High-Fidelity Synthetic Fetal MRI Image Generation

A small number of well annotated, high resolution fetal MRI datasets represent a serious barrier to the evolution of AI models applicable to the prenatal diagnosis sphere. This paper suggests an approach that uses Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) to generate synthetic images that overcome this shortcoming. The gradient-penalty loss used in WGAN-GP makes the training more stable and allows improved image fidelity. Random noise vectors are transformed into 256 x 256 x 3 synthetic MRI slices of anatomically realistic fetal brains by our generator network. The training was conducted on a dataset of fetal MRI slices with spectral normalization, AdamW optimizer, and a 5:1 critic-generator in training ratio. The measure of the Structural Similarity Index (SSIM) and manual assessment by clinical experts indicate that the created images are of high diversity and anatomical accuracy, outperforming standard GANs and DCGAN baselines. The findings provided by this paper show that WGAN-GP may be utilized in augmenting datasets used to train diagnostic models. The paper also presents the limitations and future perspectives such as use of the attention-based and conditional GANs. The proposed work will create a privacy-saving pipeline to create synthetic fetal MRIs and solve major concerns associated with scarcity or diversity of images.