Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Physics-Informed autoencoder for DSC-MRI Perfusion post-processing: application to glioma grading

Pierre Fayolle, Alexandre Bône, Noëlie Debs, Mathieu Naudin, Pascal Bourdon, Remy Guillevin, David Helbert

TL;DR

This work addresses the sensitivity of DSC-MRI perfusion analysis to noise by proposing a physics-informed autoencoder (PHAE) that learns perfusion parameters from in vivo data without external reference labels. The model combines a 1D FCN encoder that outputs $MTT$ with a physics-informed decoder that uses a Lorentzian residual to enforce realistic TRF dynamics, yielding reconstructed $C(t)$ curves. In glioma grading tasks, PHAE achieves competitive discrimination (AUC up to $0.90$) compared to standard deconvolution methods while offering shorter inference times and greater robustness to noise. By removing reliance on third-party deconvolution outputs, the approach could streamline clinical DSC-MRI post-processing and improve reliability in noisy environments.

Abstract

DSC-MRI perfusion is a medical imaging technique for diagnosing and prognosing brain tumors and strokes. Its analysis relies on mathematical deconvolution, but noise or motion artifacts in a clinical environment can disrupt this process, leading to incorrect estimate of perfusion parameters. Although deep learning approaches have shown promising results, their calibration typically rely on third-party deconvolution algorithms to generate reference outputs and are bound to reproduce their limitations. To adress this problem, we propose a physics-informed autoencoder that leverages an analytical model to decode the perfusion parameters and guide the learning of the encoding network. This autoencoder is trained in a self-supervised fashion without any third-party software and its performance is evaluated on a database with glioma patients. Our method shows reliable results for glioma grading in accordance with other well-known deconvolution algorithms despite a lower computation time. It also achieved competitive performance even in the presence of high noise which is critical in a medical environment.

Physics-Informed autoencoder for DSC-MRI Perfusion post-processing: application to glioma grading

TL;DR

This work addresses the sensitivity of DSC-MRI perfusion analysis to noise by proposing a physics-informed autoencoder (PHAE) that learns perfusion parameters from in vivo data without external reference labels. The model combines a 1D FCN encoder that outputs with a physics-informed decoder that uses a Lorentzian residual to enforce realistic TRF dynamics, yielding reconstructed curves. In glioma grading tasks, PHAE achieves competitive discrimination (AUC up to ) compared to standard deconvolution methods while offering shorter inference times and greater robustness to noise. By removing reliance on third-party deconvolution outputs, the approach could streamline clinical DSC-MRI post-processing and improve reliability in noisy environments.

Abstract

DSC-MRI perfusion is a medical imaging technique for diagnosing and prognosing brain tumors and strokes. Its analysis relies on mathematical deconvolution, but noise or motion artifacts in a clinical environment can disrupt this process, leading to incorrect estimate of perfusion parameters. Although deep learning approaches have shown promising results, their calibration typically rely on third-party deconvolution algorithms to generate reference outputs and are bound to reproduce their limitations. To adress this problem, we propose a physics-informed autoencoder that leverages an analytical model to decode the perfusion parameters and guide the learning of the encoding network. This autoencoder is trained in a self-supervised fashion without any third-party software and its performance is evaluated on a database with glioma patients. Our method shows reliable results for glioma grading in accordance with other well-known deconvolution algorithms despite a lower computation time. It also achieved competitive performance even in the presence of high noise which is critical in a medical environment.
Paper Structure (20 sections, 6 equations, 5 figures, 1 table)

This paper contains 20 sections, 6 equations, 5 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methods
  3. Data
  4. Public dataset
  5. Private dataset
  6. Data split and preprocessing
  7. Physics-informed autoencoder (PHAE)
  8. Deep encoding network
  9. Physics-informed decoder
  10. Training details
  11. Experimental setup
  12. Results
  13. Qualitative results
  14. Glioma grading performance
  15. Robustness to noise evaluation
  16. ...and 5 more sections

Figures (5)

  • Figure 1: Workflow of the PHAE. $C(t)$: tissue concentration, $C_a(t)$: arterial concentration, $FCN$: Fully Convolutional Network, CBV: Cerebral Blood Volume, CBF: Cerebral Blood Flow, MTT: Mean Transit Time, TRF: Tissue Response Function. CBV is calculated from Eq. \ref{['eq:cbv']}, CBF from Eq. \ref{['eq:cbf']}, TRF from Eqs. \ref{['eq:rf']}-\ref{['eq:trf']}, $\hat{C}(t)$ from Eq. \ref{['eq:conv']}.
  • Figure 2: Architecture of the Fully Convolutional Network (FCN). Values represent the layer parameters.
  • Figure 3: Comparison of CBF maps (grade 2 glioma) and MTT maps (grade 4 glioma) generated by oSVD wu2003tracer, Tikhonov calamante2003quantification and our methods.
  • Figure 4: Boxplots for each deconvolution methods based on CBF ratio to distinguish Low Grade Glioma (LGG) from High Grade Glioma (HGG). Each point represents a subject. The dashed lines represent the optimal cut-off values for the respective deconvolution methods.
  • Figure 5: AUC values with their confidence intervals as a function of SNR estimated for each deconvolution method.