Cross-Modality Synthesis of T1c MRI from Non-contrast Images Using GANs: Implications for Brain Tumor Research

Magnetic Resonance Imaging (MRI) has revolutionized medical diagnostics by providing a multifaceted view of the human body’s intricate structures and pathologies. With its different sequences, MRI offers comprehensive information about the structure and function of the brain, providing valuable clues in neuro-oncology. However, post-contrast (gadolinium) T1 sequences (T1c) MRI, which offers unparalleled detail in blood-brain barrier disruption, is sometimes inaccessible due to logistical and clinical constraints (e.g., costs, pregnancy, allergic reaction to the agent, etc.). Our study aims to address this issue by synthesizing T1c MRI images using generative AI methodologies. We have expanded on existing Generative Adversarial Network (GAN) frameworks and addressed the vanishing gradient problem, mode collapse, and overfitting by integrating an adaptive loss function. Our approach has shown notable improvements in image fidelity, as evidenced by the Structural Similarity Index Measure (SSIM), Peak Signal to Noise Ratio (PSNR) metrics, and visual analysis. With our conditional GAN model, we have made significant progress in medical imaging, ensuring cross-modality image synthesizing and advancing diagnostic processes in scenarios where T1c MRI access is limited.