fastMRI: Region-Aware Image Reconstruction
Improving accelerated MRI reconstruction quality by applying a region-weighted loss function to deep learning models.
Image Similarity (SSIM)
Before
0.7653
After
0.7970
Higher is Better improvement
Pixel Error (NMSE)
Before
0.0078
After
0.0075
Lower is Better improvement
Signal Quality (PSNR)
Before
31.8660
After
32.1300
Higher is Better improvement
Project Overview
The foundational fastMRI project has already had a massive real-world impact, cutting scans that once took over 30 minutes down to less than 5. My work builds on this success, tackling the core challenge: how to reconstruct highly accurate images from severely undersampled k-space data.
To solve this, I enhanced Meta's baseline U-Net (a Convolutional Neural Network) by implementing a custom 'ROI-weighted' loss function. This upgrade strategically directs the model's focus to the clinically important Region-of-Interest (ROI), improving reconstruction accuracy where it matters most. I validated the approach first with a simple, centered ROI, then evolved it using a Sobel filter for more dynamic edge-based detection.
The foundational success set the stage for the main goal: applying the same logic to a state-of-the-art Cold Diffusion model. By enhancing this powerful algorithm with the ROI-weighted technique, we achieved high-impact results, demonstrating a significant boost in image quality across all key metrics.
The ROI-weighted model (right) reconstructs the image with visibly higher detail and accuracy, outperforming the baseline model (center) and closely matching the original ground truth (left).