Reducing MRI Acquisition Times: A Deep Learning Solution (ScattNet-MR)

Date: September 03, 2025

Overview

This talk was presented at the Congreso Internacional de Supercómputo UNAM 2025, where I introduced ScattNet-MR, a deep-learning-based framework designed to reduce MRI acquisition times by up to 75% while preserving diagnostic integrity.

The method integrates super-resolution, statistical physics, and the Wavelet Scattering Transform (WST) to create a robust, interpretable, and computationally efficient reconstruction model.

Motivation

Magnetic Resonance Imaging (MRI) is one of the most powerful non-invasive diagnostic tools, yet it suffers from:

Long scan times (≈20 minutes per study)
High operational costs
Very low scanner availability in Mexico (<3 scanners per million inhabitants)

This leads to long waiting times and delayed diagnoses, especially in resource-limited healthcare systems.
(Slides 1–3)

Reducing acquisition time is therefore critical for patient throughput, clinical efficiency, and timely diagnosis.

MRI Acquisition and the Bottleneck

The talk includes an accessible overview of:

Proton precession and Larmor frequency
k-space encoding via gradient fields
Image formation through the inverse Fourier Transform
(Slides 6–8)

The key limitation:

Fully sampling k-space requires many repeated measurements, creating the inherent time bottleneck.

Acceleration via Subsampling (and Its Problems)

Reducing the number of phase-encoding lines leads to:

Aliasing
Violations of the Nyquist sampling theorem
Severe reconstruction artifacts
(Slide 10)

Deep Learning for Reconstruction

Classical interpolation fails to recover lost high-frequency structure. Deep learning offers a nonlinear approach, but…
(Slides 11–17)

Challenges of deep models in clinical reconstruction:

Risk of hallucinations (anatomically realistic but false details)
Global metrics (SSIM, PSNR) cannot guarantee structural correctness
Need for uncertainty estimation
Task-based evaluation by radiologists is essential

Our Approach: ScattNet-MR

ScattNet-MR is a residual CNN with attention mechanisms, trained not only on pixel-wise error but also using statistical constraints derived from the Wavelet Scattering Transform.
(Slides 18–24)

Key features:

WST provides stable, multiscale, and interpretable feature descriptors
KL divergence between WST coefficient distributions penalizes unrealistic reconstructions
No need to reconstruct phases explicitly
Increased robustness and reduced hallucinations

Dataset and Experimental Setup

Using the fastMRI dataset, we simulated realistic subsampling at:

50% acquisition (8% center lines)
25% acquisition (4% center lines)
(Slides 19–20)

Training details:

5000 MRI slices
14 hours on an NVIDIA RTX 4090
3.3 million trainable parameters
(Slide 25)

Results

Acceleration ×2

ScattNet-MR achieves:

High structural preservation
Improved SSIM vs. classical interpolation
(Slides 27–29)

Acceleration ×4

More challenging regime, yet the model still reconstructs realistic anatomical detail.
(Slides 30–32)

Noise robustness

When adding Gaussian noise, ScattNet-MR preserves structural consistency significantly better than CNN baselines.
(Slides 33–41)

Comparisons

Outperforms simple CNN models
Slightly below GAN models in SSIM, but with far fewer parameters and higher interpretability
(Slides 42–43)

Conclusions

Clinical relevance: Reducing MRI acquisition time directly improves patient care and hospital throughput.
Methodological innovation: Wavelet-based statistical constraints mitigate hallucinations and enhance stability.
Computational efficiency: Enables Monte Carlo uncertainty quantification in practical time.
(Slide 45)

Diego Villalba