An imageless magnetic resonance framework for fast and cost-effective decision-making

TL;DR

This work reframes MR data analysis from image reconstruction to direct interpretation of time-domain MR signals, enabling fast, low-cost decision-making with imageless MR by using optimized, low-resource sequences and pattern-recognition. In a simulated MS-lesion scenario, two fast acquisition modes (single-spoke and gradientless) were evaluated with 1D CNNs and physics-informed baselines (ART/DE), achieving high discrimination and volume-estimation performance ($AUC \approx 0.95$, $R^2 \approx 0.99$ for ideal data) and demonstrating robustness to noise and relaxation-time variability. The results suggest IMRD can enable rapid screening and monitoring in resource-limited settings, though real-world validation across modalities and lesions is needed. The framework emphasizes task-specific sequence design, minimal hardware, and a direct diagnostic signal, potentially broadening MR accessibility for triage and point-of-care use.

Abstract

Magnetic Resonance Imaging (MRI) is the gold standard in countless diagnostic procedures, yet hardware complexity, long scans, and cost preclude rapid screening and point-of-care use. We introduce Imageless Magnetic Resonance Diagnosis (IMRD), a framework that bypasses k-space sampling and image reconstruction by analyzing raw one-dimensional MR signals. We identify potentially impactful embodiments where IMRD requires only optimized pulse sequences for time-domain contrast, minimal low-field hardware, and pattern recognition algorithms to answer clinical closed queries and quantify lesion burden. As a proof of concept, we simulate multiple sclerosis lesions in silico within brain phantoms and deploy two extremely fast protocols (approximately 3 s), with and without spatial information. A 1D convolutional neural network achieves AUC close to 0.95 for lesion detection and R2 close to 0.99 for volume estimation. We also perform robustness tests under reduced signal-to-noise ratio, partial signal omission, and relaxation-time variability. By reframing MR signals as direct diagnostic metrics, IMRD paves the way for fast, low-cost MR screening and monitoring in resource-limited environments.

Figures (4)

• Figure 1: Illustration of simulated slices indicating the reference axes. Brighter areas correspond to higher $T_1$ tissues (as CSF), whereas darker tones correspond to low $T_1$ values. The MS lesion is contoured in red.a) Portion of an original phantom (i.e., before including simulated MS lesions) in the test set, showing the correspondence of axis with the image dimensions. The gradient at a 0$^{\circ}$ spoke was applied along the X-axis, the read-out direction. Slices are stacked along the Z-axis. b) Image of a healthy slice in the test set. The gradient was applied on the X-axis. c) Image of a slice with a simulated MS lesion close to the average value ($vol_\text{MS} = 1.05$ mL).
• Figure 2: Comparison of MR signals in the single-spoke and in the gradientless setups, showing the first three TRs.a) Diagram of the MR sequence for three TRs. b) First three TRs of the MR signal for single-spoke acquisition with spatial encoding by a single gradient applied in a 0$^{\circ}$ spoke along the X-axis. Solid and dashed lines show, respectively, the real (in-phase) and imaginary (quadrature) components of the complex signal induced by the transverse magnetization decay. c) MR signal for the gradientless acquisition, without spatial information, across three TRs. In this case, the FID registered at $B_0$ (with no gradients applied) is observed after each resonant RF pulse, featuring exponential decays that hold the discriminant information between MS and healthy slices. The horizontal axis represents time points at which the global transverse magnetization signal is registered, losing the spatial information encoded in the frequency gradient from single-spoke.
• Figure 3: Validation of lesion volume estimation for single-spoke and gradientless data.a) Regression results for single-spoke acquisition, with predicted ($\hat{vol}_{\text{MS}}$) versus simulated lesion volumes ($vol_{\text{MS}}$) showing increased prediction variability for $vol_{\text{MS}}>0$. b) Example of a brain slice with black pixels representing a simulated MS lesion volume $vol_\text{MS}$ of $\approx 0.06$ mL, the maximum undetected volume (False Negative) for the single-spoke model. c) Reconstructed image from first TR for the slice with a simulated MS volume $\approx 0.06$ mL, which is the maximum undetected volume with the single-spoke setup. d) Regression results for a gradientless acquisition, with predicted ($\hat{vol}_{\text{MS}}$) versus simulated lesion volumes ($vol_\text{MS}$) showing high agreement with an $R^2$ and a slope close to unity. e) Example of a brain slice with black pixels representing a simulated MS lesion volume $vol_{\text{MS}} \approx 0.26$ mL, the maximum undetected volume (False Negative) for the gradientless model. f) Example of a brain slice with black pixels representing a simulated MS lesion volume $vol_{\text{MS}} \approx 0.03$ mL, the second maximum undetected volume (False Negative) for the gradientless model.
• Figure 4: Performance of CNNs for MS lesion volume estimation and detection under varying SNR and apodization levels. The left axis scale corresponds to model performance metrics ($AUC$ and $R^2$, black circles and square markers, respectively). The right axis scale corresponds to mL of either simulated MS volume (for max. $vol_\text{MS}^\text{FN}$) or predicted MS volume (for max. $\hat{vol}_\text{MS}^\text{FP}$) a)Single-spoke performance metrics as a function of SNR. b)Single-spoke performance metrics as a function of the high-frequency information included, which is equivalent to considering regions with a higher applied gradient. c)Gradientless performance metrics as a function of SNR. d)Gradientless performance metrics as a function of the apodization level applied to the FID signal, losing last instants, when the magnetization vector is already stable.