Reducing MRI Acquisition Times with Deep Learning (ScattNet-MR)
Date:
Overview
At MEXCAS 2025, I presented the same work delivered at the
Congreso Internacional de Supercómputo UNAM 2025:
a deep-learning framework called ScattNet-MR, designed to significantly reduce MRI acquisition times while preserving diagnostic fidelity.
This project integrates:
- Super-resolution techniques,
- Statistical physics,
- Wavelet Scattering Transform (WST), and
- A residual convolutional architecture with attention,
to reconstruct high-quality MRI images from heavily undersampled k-space data.
The method achieves up to 75% reduction in scan time without degrading critical anatomical information.
Motivation
Modern MRI presents a critical bottleneck:
- Long acquisition times (~20 minutes per study)
- High equipment cost
- Limited access in Mexico (<3 scanners per million inhabitants)
This results in delays for diagnosis and treatment.
ScattNet-MR aims to address these systemic constraints by accelerating acquisition without sacrificing medical reliability.
Method Summary
ScattNet-MR reconstructs high-quality MRI images by combining:
Wavelet Scattering Transform (WST)
A mathematically stable, multi-scale representation that:
- Captures non-Gaussian and structural information
- Is robust to local deformations
- Provides a physically interpretable loss function
Residual CNN with attention
Learns the mapping from undersampled MRI slices to full-resolution images using:
- Multi-channel feature fusion
- Statistical consistency enforced by WST-based KL divergence
- Regularization against hallucinations common in GAN-based MRI models
Dataset & Training
The model was trained on the fastMRI dataset (Facebook AI Research + NYU Langone Health), using:
- 5000 MRI slices
- Acceleration factors of 2× and 4×
- 14 hours of training on an NVIDIA RTX 4090
- 3.3M trainable parameters
Results
- High structural similarity (SSIM) for both 2× and 4× acceleration
- Strong robustness to Gaussian acquisition noise
- Superior stability compared to CNN-only baselines
- Competitive performance with GAN models while using far fewer parameters and offering better interpretability
ScattNet-MR enables complex statistical experiments (e.g., Monte-Carlo dropout) to run efficiently on consumer-grade GPUs.
Clinical Potential
ScattNet-MR has direct impact on:
- Patient throughput
- Workflow efficiency
- Early disease detection
- Computationally efficient MRI-based research
Its physics-aware design reduces the risk of hallucinated structures, supporting safer medical deployment.