Fast & Sustainable Brain MRI

The MRI sustainability problem

Lengthy MRI acquisition negatively impacts sustainability through high energy consumption, increased patient discomfort, and reduced clinical efficiency. MRI is the single largest energy consumer in radiology, and its environmental footprint is growing on both fronts — more scans and more installed units.

Reducing acquisition duration directly addresses this: shorter sequences mean fewer gradient pulses, fewer RF cycles, and less cooling overhead — decreasing per-patient energy consumption, minimizing environmental impact, and improving scanner throughput.

The limits of single-contrast reconstruction

Traditional acceleration techniques such as parallel imaging and compressed sensing successfully reduce acquisition times, but introduce reconstruction artifacts, offer limited gains at higher accelerations, and add computational demands. Recovering a scan from a single aggressively undersampled contrast is inherently ill-posed — auxiliary information is essential.

Prior multi-modal methods typically reconstruct accelerated contrasts using a reference contrast acquired at high resolution. We remove that dependency: we exploit complementary information from triplets of scans with mutually orthogonal phase-encoding directions, with no high-resolution priors and no vendor dependency.

Our approach: orthogonal triple-modality undersampling

Acceleration is achieved by selectively reducing high-frequency components — undersampling roughly 3/4 of k-space along a distinct phase-encoding direction for each of the three contrasts (T1, T2, FLAIR). This shortens phase-encoding and readout durations, preserves essential low-frequency image contrast, reduces susceptibility artifacts, and substantially cuts gradient and RF energy consumption.

Because the blur direction differs per contrast, the missing high-frequency detail in one modality is captured by another. To exploit this, we first register the different contrasts to one another and then to an MNI atlas — improved spatial alignment significantly enhances network performance. Reconstruction then uses a custom deblurring network trained with a combined SSIM + L1 perceptual loss derived from VGG-19 features.

The reconstruction engine: FARD

Reconstruction is carried out by FARD — Frequency Attention Residual Denoising, a network designed for multi-contrast MRI deblurring with anisotropic kernels elongated in the direction of blur. FARD combines processing in both the frequency and image domains (global + local), and enforces sparsity in the attention domain — inspired by compressed sensing — so that attention favors anatomical structure over noise.

Results

The method was validated on healthy subjects scanned at 1.5T (Philips Ingenia), each undergoing both the accelerated protocol and clinical high-resolution scans as reference, and further validated on a large synthetically subsampled dataset (BRATS). Visual inspection consistently shows restoration of fine detail despite aggressive acceleration — and image quality improves monotonically as more complementary contrasts are added.

Quantitatively, adding contrasts consistently raises both SSIM and PSNR for every target modality:

Acceleration reduces power consumption

The acceleration translates directly into sustainability gains on the scanner. For the full T1 + T2 + FLAIR protocol on a Philips 1.5T Ingenia, total acquisition drops from 10.4 minutes to 2.61 minutes, and total energy from 2.93 kWh to 0.72 kWh — up to a 75% reduction in energy per sequence.

Summary & outlook

The triple-modality AI-deblurring method significantly contributes to MRI sustainability by reducing acquisition times and energy consumption while maintaining diagnostic image quality. Registration across contrasts and to a standard MNI atlas enhances reconstruction accuracy and robustness, and because the reconstruction is scanner-independent, the approach offers broad applicability.

Initial validation on healthy subjects confirms feasibility and diagnostic efficacy. Future work will extend the evaluation to pathological conditions and address residual artifacts such as potential Gibbs ringing — with the goal of enabling sustainable, high-throughput imaging in high-volume or resource-constrained clinical settings.

References

1. Hod, G., Green, M., Waserman, M., Konen, E., Shrot, S., Nelkenbaum, I., Kiryati, N. and Mayer, A., 2022. Complementary phase encoding for pair-wise neural deblurring of accelerated brain MRI. In Computer Vision – ECCV 2022 Workshops.
2. Xiang, L., Chen, Y., Chang, W., Zhan, Y., Lin, W., Wang, Q. and Shen, D., 2019. Deep-learning-based multi-modal fusion for fast MR reconstruction. IEEE Transactions on Biomedical Engineering, 66(7), pp.2105–2114.
3. Maintz, J.B.A. and Viergever, M.A., 1998. A survey of medical image registration. Medical Image Analysis, 2(1), pp.1–36.
4. Mayberg, M. et al., 2022. Anisotropic neural deblurring for MRI acceleration. International Journal of Computer Assisted Radiology and Surgery, 17, pp.315–327.
5. Feng, C.-M. et al., 2023. Multi-contrast MRI reconstruction. IEEE Transactions on Medical Imaging, 42(10), pp.2804–2816.

Content on this page is drawn from the ISMRM 2025 submission “Fast and Sustainable Brain MRI by Triple-Modality Deblurring” (A. Nazarov, D. Roizen, G. Greenberg, N. Kiryati, A. Mayer) and the accompanying project presentation.