[Paper Review] State estimation with nonlinear reduced models. Application to the reconstruction of blood flows with Doppler ultrasound images

Felipe Galarce, Jean-Frédéric Gerbeau+2 more|arXiv (Cornell University)|Dec 10, 2019

Electrical and Bioimpedance Tomography5 citations

TL;DR

This paper proposes a nonlinear state estimation framework using task-adapted reduced-order models for fast, accurate 3D reconstruction of blood flow from Doppler ultrasound images. By constructing reduced spaces tailored to the reconstruction task—via manifold partitioning and local POD bases—it achieves up to tenfold improvement in accuracy over classical linear PBDW methods, particularly excelling in blockage detection and flow ratio estimation with conservative overestimation to prevent false negatives.

ABSTRACT

Over the past years, several fast reconstruction algorithms based on reduced models have been proposed to address the state estimation problem of approximating an unknown function u of a Hilbert space V from measurement observations. Most strategies are however based on linear mappings where the reduced model is built a priori and independently of the observation space and the measurements. In this work we explore some nonlinear extensions that take these elements into account in the construction of the basis. The methodology is applied to the reconstruction of 3D blood flows from Doppler ultrasound images. The example not only shows the good performance of the nonlinear methods, but it also illustrates the potential of the methodology in medicine as tool to process and interpret data in a systematic manner which could help to build more robust and individualized diagnostics.

Motivation & Objective

• To improve the accuracy and speed of 3D blood flow reconstruction from partial Doppler ultrasound measurements in clinical settings.
• To address limitations of classical linear reduced-order models that are built a priori and independently of the reconstruction task.
• To develop and evaluate nonlinear reconstruction methods that adapt the reduced model space to the measurement data and task-specific objectives.
• To assess the performance of data-driven and locally adapted reduced models in comparison to standard Proper Orthogonal Decomposition (POD) and PBDW approaches.
• To evaluate the method’s reliability for clinical decision support, particularly in detecting arterial blockages via flow ratio estimation.

Proposed method

• The study employs a modified Parametrized Background Data-Weak (PBDW) framework, where the reduced model space is no longer built a priori but adapted to the reconstruction task.
• It introduces two novel reduced space constructions: (1) manifold partitioning with local POD bases on each partition, and (2) data-driven reduced spaces derived directly from Doppler measurements.
• The reconstruction is formulated as a least-squares fitting problem between measurements and the reduced model, augmented with a correction term for model bias.
• The method is tested on synthetic, semi-realistic Doppler ultrasound data simulating peak systole flow in carotid arteries with and without stenosis.
• Reconstruction quality is evaluated using error metrics (mean and worst-case) and flow ratio accuracy for blockage detection.
• The approach is compared across modalities: Color Flow Imaging (CFI) and Vector Flow Imaging (VFI), with VFI showing slightly better reconstruction quality.

Experimental results

Research questions

• RQ1 Can task-adapted reduced-order models significantly improve the accuracy of 3D blood flow reconstruction from partial Doppler ultrasound measurements?
• RQ2 How does the performance of nonlinear reconstruction methods—based on local POD or data-driven spaces—compare to classical linear PBDW with global POD?
• RQ3 To what extent do measurement-dependent reduced spaces enhance reconstruction fidelity, especially in low-signal regions like diastole?
• RQ4 Can the proposed method reliably estimate hemodynamic indices such as flow ratio for arterial blockage detection?
• RQ5 Is the reconstruction robust and conservative enough to prevent false-negative diagnoses in clinical applications?

Key findings

• The nonlinear reconstruction method based on manifold partitioning and local POD bases outperforms all other methods, achieving up to tenfold improvement in reconstruction accuracy over classical linear PBDW with global POD.
• The method delivers highly accurate flow ratio estimation for blockage detection, with a threshold of r* = 1.25 distinguishing healthy from sick patients.
• For r > 1.7, the method overestimates the flow ratio, which is clinically preferable to underestimation, ensuring no false-negative diagnoses.
• The VFI modality yields slightly better reconstruction quality than CFI, though both provide satisfactory predictions across all test cases.
• The data-driven reduced model approach performed worse than local POD but still outperformed classical PBDW, suggesting limitations in learning from sparse or low-signal measurements, especially in diastole.
• The proposed framework enables fast, reliable, and safe 3D flow reconstruction, making it suitable for real-time clinical decision support in hemodynamics.