SUPER-IVIM-DC-BOOT: A Bootstrapped Physics-Informed IVIM Framework for Robust Placental Microstructure Analysis in Uncontrolled Maternal Diabetes

Naama Gavrielov, Moran Gawie-Rotman, Daphna Link-Sourani, Abdel-Rauf Zeina, Roni Shreter, Esther Maor-Sagie, Rinat Gabbay-Benziv, Moti Freiman

Keywords: Physics-Informed Deep Learning, Diffusion-Weighted MRI, IVIM (Intra-voxel Incoherent Motion), bootstrap resampling, Maternal Diabetes, Placental Functional Assessment, Placental Microstructure

TL;DR: SUPER-IVIM-DC-BOOT, a physics-informed IVIM network with bootstrap resampling, stabilizes low-SNR placental MRI and detects microstructural abnormalities in uncontrolled maternal diabetes that conventional methods miss.

Abstract: Intravoxel Incoherent Motion (IVIM) modeling decomposes diffusion-weighted MRI (DWI) signals into diffusion- and perfusion-related components, enabling non-invasive characterization of microvascular structure in highly vascularized tissues such as the placenta. However, accurate recovery of IVIM parameters remains an ill-posed inverse problem, particularly under the low signal-to-noise ratio (SNR) and sparse b-value sampling common in fetal and placental imaging. We introduce SUPER-IVIM-DC-BOOT, an acquisition-aware, physics-informed neural network that integrates explicit b-value encoding with inference-time bootstrap resampling to stabilize bi-exponential parameter estimation. By aggregating predictions from stratified subsampled inputs and enforcing a data-consistent forward-model constraint, the method improves robustness to noise and protocol sparsity. Numerical simulations demonstrate that bootstrap aggregation substantially reduces Normalized Root Mean Squared Error (NRMSE) in low-SNR regimes (SNR 7–10) relative to non-bootstrapped baselines. In a healthy volunteer study, SUPER-IVIM-DC-BOOT recovered IVIM parameters from a sparse 9 b-value protocol with accuracy comparable to reference estimates obtained from a dense 16 b-value acquisition. Applied to placental DWI in pregnancies with uncontrolled maternal diabetes ($N=6$) versus matched controls ($N=8$), the method detected significant alterations in perfusion fraction (f) and tissue diffusivity (D) ($p<0.05$), whereas conventional Trust-Region Reflective fitting (SLS-TRF) underestimated perfusion changes and failed to identify restricted diffusivity. These findings demonstrate that combining protocol-aware physics constraints with inference-time bootstrapping yields robust and clinically meaningful IVIM quantification under realistic, sparse acquisition conditions. Code will be released upon acceptance.

Primary Subject Area: Image Acquisition and Reconstruction

Secondary Subject Area: Application: Radiology

Registration Requirement: Yes

Reproducibility: https://github.com/naamagav/SUPER-IVIM-DC-BOOT

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 87