Generative MRI Restoration

The Challenge

Magnetic Resonance Imaging (MRI) diagnostic quality is often bottlenecked by hardware limitations, specifically Rician Noise and Bias Fields (intensity inhomogeneities).

Standard solutions like N4ITK rely on slow, iterative CPU algorithms to correct these errors one by one. I investigated if Generative AI could act as a “software upgrade” for older scanners, solving both problems instantly.

The Solution

I engineered a Latent Diffusion Pipeline that treats MRI correction as a single-pass generative task rather than an iterative optimization problem.

• N4-Guided Conditioning: Engineered a conditioning signal that guides the diffusion model to “see” the clean anatomy through the noise, ensuring medical accuracy is preserved.
• Modified Noise Scheduler: Implemented a custom noise envelope that prevents the model from hallucinating or blurring fine anatomical details—a common failure mode in standard Stable Diffusion.
• Physics-Based Training: Generated synthetic training data that simulates the actual RF coil physics of MRI scanners, ensuring the model works on real-world hardware artifacts.

The Impact

• Quality: Achieved a 68% reduction in structural error (SSIM) compared to industry-standard baselines, effectively rendering the images “visually identical” to ground truth.
• Efficiency: Converted a multi-step iterative correction process into a single-shot inference pass, enabling near real-time image enhancement.
• Business Value: Demonstrated that software-defined correction can compensate for hardware imperfections, potentially reducing the manufacturing cost of MRI receiver coils.

---

Note: Experiments and implementations for this project are scattered across multiple repositories:

• MRI_Denoising - Main repository (2D/3D implementations)
• nn_MONAI - MONAI-based 3D UNet experiments for volumetric denoising
• DDM2 - Self-supervised diffusion model for MRI denoising (custom modifications)
• SupervisedDiffusion - Supervised diffusion models for paired noisy-clean MRI data
• MRI_Denoising_Old - Legacy PyTorch Lightning implementation with UNet architectures