Decoding the human brain tissue response to radiofrequency excitation using a biophysical-model-free deep MRI on a chip framework

TL;DR

This work addresses the time burden of multi-contrast MRI by introducing DeepMonC, a biophysical-model-free framework that decodes brain tissue response to RF excitation and generates on-demand molecular and quantitative contrasts from a 28.2 s calibration scan. A vision-transformer-based core module translates calibration and RF information into spatiotemporal spin dynamics, while a transfer-learning quantification module outputs six tissue- and scanner-related parameter maps (k_ssw, f_ss, B0, B1, T1, T2). The approach achieves high-fidelity reconstructions (SSIM > 0.96, PSNR > 36) and delivers a 94% reduction in scan time compared with conventional protocols, generalizing across unseen subjects, pathologies, and scanner models. This framework holds promise to accelerate clinical MRI by delivering rich molecular and biophysical information in substantially shorter exams and could extend to other organs or modalities with appropriate training.

Abstract

Magnetic resonance imaging (MRI) relies on radiofrequency (RF) excitation of proton spin. Clinical diagnosis requires a comprehensive collation of biophysical data via multiple MRI contrasts, acquired using a series of RF sequences that lead to lengthy examinations. Here, we developed a vision transformer-based framework that captures the spatiotemporal magnetic signal evolution and decodes the brain tissue response to RF excitation, constituting an MRI on a chip. Following a per-subject rapid calibration scan (28.2 s), a wide variety of image contrasts including fully quantitative molecular, water relaxation, and magnetic field maps can be generated automatically. The method was validated across healthy subjects and a cancer patient in two different imaging sites, and proved to be 94% faster than alternative protocols. The deep MRI on a chip (DeepMonC) framework may reveal the molecular composition of the human brain tissue in a wide range of pathologies, while offering clinically attractive scan times.

Figures (6)

• Figure 1: Schematic representation of the biophysical-model-free deep MRI on a chip (DeepMonC) framework. a. Automatic prediction of unseen molecular MRI contrast weighted images. A multi-domain input is used, including a sequence of m non-steady-state MRI calibration images and an RF excitation parameter tensor. It includes the acquisition parameters associated with the calibration images (solid lines) and the on-demand acquisition parameters (dashed lines) for the desired image output (m new images shown at the top). Separate embeddings for the real image data and the physical RF properties are learned using a vision transformer and a fully connected layer, respectively. b. A quantification module for the simultaneous mapping of six tissue and scanner parameter maps, including the semi-solid proton volume fraction (f$_{ss}$) and exchange rate (k$_{ssw}$), water proton longitudinal (T$_1$) and transverse (T$_2$) relaxation, and static (B$_0$) and transmit (B$_1$) magnetic fields. This module exploits the multi-domain embedding learned by the core module, utilizing a transfer learning strategy.
• Figure 1: Performance analysis for on-demand generation of molecular contrast-weighted images, comparing the DeepMonC reconstructed output to the reference ground truth. SSIM - Structural similarity index measure; PSNR - peak signal-to-noise ratio; NRMSE - normalized mean-square error.
• Figure 2: Automatic prediction of unseen molecular MRI contrast weighted images. A comparison between representative ground truth (a, c, e) and DeepMonC-predicted (b, d, f) molecular MRI contrast-weighted images in the human brain. (a, b) Semiolid MT-weighted images from an unseen subject. (c, d) Amide proton transfer CEST-weighted images from a brain tumor patient scanned at an unseen imaging site. (e, f) Semisolid MT-weighted images from an unseen subject scanned at an unseen imaging site with hardware that was different from that used for training.
• Figure 3: Quantitative reconstruction of six molecular MRI, scanner field, and water-proton relaxation quantitative maps from a new healthy human volunteer scanned at the same imaging site used for training. (a) Ground truth reference images obtained using conventional T$_1$ and T$_2$-mapping, WASABI, and semisolid MT MR-Fingerprinting (MRF) in 8.5 min. (b) The same parameter maps obtained using DeepMonC in merely 28.2 s (94% scan time acceleration). Note the reduced field inhomogeneity (as seen in the B$_0$ and B$_1$ predicted images), which explains the successful noise reduction in the output maps (white arrows). (c) Quantitative reconstruction using conventional supervised learning (RF tissue response pretraining excluded), utilizing the same raw input data used in (b) for comparison. (d) Statistical analysis of the SSIM, PSNR, and NRMSE performance measures, comparing the DeepMonC reconstructed parameter maps to reference ground truth (n = 69 brain image slices per group ). ****p$<$0.0001.
• Figure 4: Quantitative reconstruction of six molecular MRI, scanner field, and water-proton relaxation quantitative maps from a brain cancer patient scanned at a different imaging site compared to training. (a) Ground truth reference images obtained using conventional T$_1$ and T$_2$-mapping, WASABI, and semisolid MT MR-Fingerprinting (MRF) in 8.5 min. (b) The same parameter maps obtained using DeepMonC in merely 28.2 s (94% scan time acceleration). (c) Quantitative reconstruction using conventional supervised learning (RF tissue response pretraining excluded), utilizing the same raw input data used in (b) for comparison. (d). Statistical analysis of the SSIM, PSNR, and NRMSE performance measures, comparing the DeepMonC reconstructed parameter maps to reference ground truth (n = 68 brain image slices per group ). ****p$<$0.0001.