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

Estimating multi-component $T_{2}$ relaxation distribution from Multi-Echo Spin Echo (MESE) MRI is a severely ill-posed inverse problem, traditionally approached with regularized non-negative least squares (NNLS). In abdominal imaging, and in the pancreas in particular, low Signal-to-Noise Ratio (SNR), and residual uncorrelated noise between reconstructed echoes challenge both classical solvers and deterministic deep learning models. We introduce a bootstrap-based inference framework for robust distributional $T_2$ estimation, which performs stochastic resampling of the echo train and aggregates predictions across multiple echo subsets. This strategy treats the acquisition as a distribution rather than a fixed input, yielding variance-reduced, physically consistent estimates and converting deterministic relaxometry networks into probabilistic ensemble predictors. Building on the P2T2 architecture, our method applies inference-time bootstrapping to smooth residual noise artifacts, increase tolerance to stochastic inference errors, and enhance fidelity to the underlying relaxation distribution. We demonstrate a clinical application of the proposed approach for functional and physical assessment of the pancreas. Currently available techniques for noninvasive pancreatic evaluation are limited due to the organ’s concealed retroperitoneal location and the procedural risks associated with biopsy, driven in part by the high concentration of proteases that can leak and cause intra-abdominal infection. These constraints highlight the need for functional imaging biomarkers capable of capturing early pathophysiological changes. A prominent example is type 1 diabetes (T1DM), in which progressive destruction of beta cells begins years before overt hyperglycemia, yet no existing imaging modality can assess early inflammation or the decline of pancreatic islets. A further unmet need lies in characterizing pancreatic lesions suspected of malignancy: although malignant and benign lesions differ in their physical properties, current imaging methods do not reliably distinguish between them. To examine the clinical utility of our method, we evaluate performance in test–retest reproducibility study ($N=7$) and a T1DM versus healthy differentiation task ($N=8$). The proposed approach achieves the lowest Wasserstein distances across repeated scans and demonstrates superior sensitivity to subtle, physiology-driven shifts in the relaxation-time distribution, outperforming classical NNLS and non-bootstrapped deep learning baselines. These results establish inference-time bootstrapping as an effective and practical enhancement for quantitative $T_2$ relaxometry in low-SNR abdominal imaging, enabling more stable and discriminative estimation of relaxation-time distributions.