Bootstrapped Physically-Primed Neural Networks for Robust T2 Distribution Estimation in Low-SNR Pancreatic MRI

Hadas Ben-Atya, Nicole Abramenkov, Noa Mashiah, Luise Brock, Daphna Link-Sourani, Ram Weiss, Moti Freiman

Keywords: Multi-component T2, Quantitative MRI, Physics-informed Neural Networks, Bootstrapped Inference, Test-Retest Reliability, Microstructural Biomarkers

TL;DR: A bootstrap-based inference method stabilizes multi-component T2 estimation in low-SNR abdominal MRI, reducing noise sensitivity and improving reproducibility and pancreatic tissue differentiation over classical and deep learning approaches.

Abstract:

Estimating multi-component T2 relaxation distributions from Multi-Echo Spin Echo (MESE) MRI is an ill-posed inverse problem, traditionally solved via regularized non-negative least squares (NNLS). In abdominal imaging, particularly of the pancreas, low Signal-to-Noise Ratio (SNR) and physiological motion render classical solvers and standard deep learning models unstable.
We introduce a bootstrap-based inference framework for stable distributional T2 estimation, which leverages stochastic resampling of the echo train to produce variance-reduced, physically consistent predictions. This approach transforms deterministic relaxometry networks into probabilistic estimators, substantially mitigating the effects of low SNR and corrupted echoes.

Building on our previous P2T2 architecture, the proposed method performs inference-time bootstrapping by aggregating predictions across random echo subsets, effectively smoothing noise-induced artifacts and enhancing distributional fidelity.
We evaluated this method on two cohorts: a test-retest reproducibility study (N=7) and a T1DM versus healthy differentiation task (N=8). The proposed method achieved the lowest Wasserstein distances across repeated scan and demonstrate superior sensitivity to subtle microstructural shifts, outperforming both baseline deep learning and classical methods.
These findings establish inference-time bootstrapping as a practical and robust enhancement for quantitative T2 relaxometry in low-SNR abdominal imaging.

Primary Subject Area: Image Acquisition and Reconstruction

Secondary Subject Area: Application: Radiology

Registration Requirement: Yes

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 110