Automated quality control of synthetic brain MRI images: a reproducible approach

Synthetic images have the potential to play an important role in training deep learning models, particularly when real data is limited or privacy concerns arise, such as in the medical field. However, validating synthetic images is necessary to ensure their fidelity and relevance to real-world scenarios. Additionally, emphasizing the need for reproducibility in scientific work ensures that findings can be independently verified, fostering trust and advancing research integrity. Incorporating these practices not only enhances the robustness of AI models but also bolsters the credibility and sustainability of scientific endeavors. The general purpose of this thesis is to develop a reproducible pipeline to automatically clean synthetic brain magnetic resonance (MR) images generated using Progressive Auxiliary Classifier GAN (PACGAN) by eliminating unrealistic data samples. The heart of the pipeline is constituted by a neural network that classifies the images into realistic/not realistic samples. This network has been trained on a dataset comprising 24 features extracted from 2000 synthetic images generated by PACGAN. These features have been selected by a set of suitable metrics commonly used to assess the quality of synthetic MR images. Our neural network reached an area under the ROC curve (AUC) of 0.74, proving that the selected features were effective in selecting synthetic realistic images. All the work has been carried out on CNAF's cloud infrastructure, using a virtual machine with a Tesla T4 GPU and 1 TB of NVMe.