Deep Learning Generative Models for Breast MRI Synthesis: A 3D-GAN Comparative Study

Cancer remains a significant contributor to global mortality, with breast cancer emerging as a prevalent concern, particularly among women. Early detection significantly improves treatment outcomes, making it imperative to enhance diagnostic processes. Magnetic Resonance Imaging (MRI) has become a pivotal tool in breast cancer imaging, offering precision and independence from breast density. However, the reliance on expert interpretation and potential delays in diagnosis pose challenges for this imaging technique. This work addresses these challenges by using machine learning, to aid in breast cancer detection. Traditional computer-aided diagnosis (CAD) systems, inspired by mammography and MRI, require extensive and diverse datasets for effective training. Yet, the scarcity of such datasets, in addition to the preservation of privacy, needs innovative approaches for augmentation. Generative models, including autoregressive models, variational autoencoders (VAEs), generative adversarial networks (GANs) and diffusion models, have emerged as valuable tools for generating synthetic medical data. Motivated by the, frequent, three-dimensional nature of medical images, this study explores the synthesis of high-resolution 3D breast MRI images using advanced generative models, such as: 3D WGAN-GP, 3D α-WGAN-GP and HA-GAN. Initially, 3D α-WGAN-GP outperformed 3D WGAN-GP at 64 resolution based on FID, MMD, and MS-SSIM metrics. HA-GAN excelled at 128 resolution, generating artifact-free images surpassing the previous models and obtaining FID, MMD, and MS-SSIM values of 189.724, 200879.92, and 0.4136, respectively. Additionally, a visual test was conducted with a healthcare professional, where the model achieved an accuracy of 71.67%. Finally, a Real-ESRGAN model, developed by a colleague (Pedro Santos Pereira), further enhanced image clarity, crucial for detailed medical diagnostics, by increasing the resolution of the images generated by HA-GAN from 128 to 512.