CD-GANs : Conditional dynamic training with GANs for enhanced post-gadolinium glioblastoma MRIs

Gliomas are brain tumors originating from glial cells, graded I to IV by the WHO, with Grade IV glioblastoma (GBM) being the most aggressive. Magnetic resonance imaging (MRI) is crucial for monitoring disease progression and guiding therapy in glioma patients. However, gadolinium, a contrast agent used for visualization, poses risks of accumulation, particularly in patients with nephrogenic systemic fibrosis, renal disease, diabetic nephropathy, or low glomerular filtration rate. This underscores the need for innovative imaging solutions. This paper proposes a method for generating synthetic T1-weighted Contrast-Enhanced (T1CE) images, enabling tumor visualization without gadolinium injection. This approach aims to improve patient outcomes by accurately delineating tumor characteristics while avoiding gadolinium-related risks. We introduce a Generative Adversarial Network (GAN) for T1CE MRI synthesis, utilizing a dynamic training mechanism with condition-based discriminators adjusted based on tumor presence in each batch. This ensures precise T1CE image synthesis. Furthermore, this work introduces progressive intensity training, which prioritizes learning from dark intensity regions before gradually incorporating medium and light intensities, enhancing fine detail capture. This strategy is applied to Pix2pix and our two frameworks, Conditional-Patch Discriminator (CPD-GAN) and Double Conditional Discriminators (DCD-GAN), resulting in their progressive intensity versions: PiPix2pix, PiCPD-GAN, and PiDCD-GAN. Validated on the Brats2020 dataset, the proposed methodology demonstrates superior performance in multi-modality imaging compared to state-of-the-art methods. Experimental results, particularly for DCD-GAN and its Pi variant, highlight their effectiveness in T1CE image synthesis, offering a safer and more reliable alternative for glioma diagnosis and monitoring.