Quantitative mapping from conventional MRI using self-supervised physics-guided deep learning: applications to a large-scale, clinically heterogeneous dataset

TL;DR

Results indicate that the proposed framework can robustly transform diverse clinical conventional MRI data into quantitative maps, potentially paving the way for large-scale quantitative biomarker research.

Abstract

Magnetic resonance imaging (MRI) is a cornerstone of clinical neuroimaging, yet conventional MRIs provide qualitative information heavily dependent on scanner hardware and acquisition settings. While quantitative MRI (qMRI) offers intrinsic tissue parameters, the requirement for specialized acquisition protocols and reconstruction algorithms restricts its availability and impedes large-scale biomarker research. This study presents a self-supervised physics-guided deep learning framework to infer quantitative T1, T2, and proton-density (PD) maps directly from widely available clinical conventional T1-weighted, T2-weighted, and FLAIR MRIs. The framework was trained and evaluated on a large-scale, clinically heterogeneous dataset comprising 4,121 scan sessions acquired at our institution over six years on four different 3 T MRI scanner systems, capturing real-world clinical variability. The framework integrates Bloch-based signal models directly into the training objective. Across more than 600 test sessions, the generated maps exhibited white matter and gray matter values consistent with literature ranges. Additionally, the generated maps showed invariance to scanner hardware and acquisition protocol groups, with inter-group coefficients of variation $\leq$ 1.1%. Subject-specific analyses demonstrated excellent voxel-wise reproducibility across scanner systems and sequence parameters, with Pearson $r$ and concordance correlation coefficients exceeding 0.82 for T1 and T2. Mean relative voxel-wise differences were low across all quantitative parameters, especially for T2 ($<$ 6%). These results indicate that the proposed framework can robustly transform diverse clinical conventional MRI data into quantitative maps, potentially paving the way for large-scale quantitative biomarker research.

Figures (14)

• Figure 1: Overview of the proposed self-supervised physics-guided CNN framework. Each input contrast (T1w, T2w, FLAIR) is processed independently by a three-level (+ bottleneck level) shared-weight encoder. At each level, an attention fusion module applies global average pooling (GAP) and a shared multilayer perceptron (MLP). Subsequently, attention weights are computed using a across contrasts. The decoder outputs three channels (T1, T2, PD maps) from the fused features. From these quantitative maps, conventional MRI is synthesized using Bloch-based physical signal models. To train the deep learning model, an L1-loss is computed between the input and synthesized images. Additionally, total variation regularization (TV reg.) is applied to all quantitative maps, along with a soft lower bound on the PD map (PD constraint).
• Figure 2: Representative example slices from the clinical archive test set. For seven subject (A--G), featuring various lesions, the input contrasts (T1w, T2w, and FLAIR) are shown alongside the corresponding generated quantitative T1, T2, and PD maps. (A) Multiple sclerosis (MS); (B) Cerebral cavernous venous malformation; (C) Meningioma; (D) Glioblastoma, status post-partial resection, chemoradiotherapy and post-radiation temozolomide; (E) High-grade glioma progression, status post-radiotherapy and chemotherapy; (F) Oligodendroglioma, status post-radiotherapy and chemotherapy; (G) Multifocal high-grade glioma, status post-radiotherapy. For patients A--C there was no ongoing treatment at the time of imaging.
• Figure 3: Summary of quantitative T1, T2, and PD values across all 603 clinical test sessions. Top: distributions of T1, T2, and PD values values in white matter (WM) and gray matter (GM). Solid lines indicate the mean value across sessions and shaded areas indicate one standard deviation. Bottom: boxplots of session-wise mean WM and GM values for T1, T2, and PD.
• Figure 4: Boxplots of session-wise mean quantitative T1, T2, and PD values in WM and GM of the test set grouped by scanner system (Achieva, Ingenia, Ingenia CX, and Ingenia Elition X). The number of scan sessions in each group is displayed in the legend.
• Figure 5: Boxplots of session-wise mean quantitative T1, T2, and PD values in WM and GM of the test set grouped by protocol cluster. The number of scan sessions in each group is displayed in the legend.