Accelerating physics-informed neural fields for fast CT perfusion analysis in acute ischemic stroke

Published: 06 Jun 2024, Last Modified: 06 Jun 2024MIDL 2024 OralEveryoneRevisionsBibTeXCC BY 4.0
Keywords: neural fields, CT perfusion, physics-informed, acute ischemic stroke
Abstract: Spatio-temporal perfusion physics-informed neural networks were introduced as a new method (SPPINN) for CT perfusion (CTP) analysis in acute ischemic stroke. SPPINN leverages physics-informed learning and neural fields to perform a robust analysis of noisy CTP data. However, SPPINN faces limitations that hinder its application in practice, namely its implementation as a slice-based (2D) method, lengthy computation times, and the lack of infarct core segmentation. To address these challenges, we introduce a new approach to accelerate physics-informed neural fields for fast, volume-based (3D), CTP analysis including infarct core segmentation: ReSPPINN. To accommodate 3D data while simultaneously reducing computation times, we integrate efficient coordinate encodings. Furthermore, to ensure even faster model convergence, we use a meta-learning strategy. In addition, we also segment the infarct core. We employ acute MRI reference standard infarct core segmentations to evaluate ReSPPINN and we compare the performance with two commercial software packages. We show that meta-learning allows for full-volume perfusion map generation in 1.2 minutes without comprising quality, compared to over 40 minutes required by SPPINN. Moreover, ReSPPINN's infarct core segmentation outperforms commercial software.
Latex Code: zip
Copyright Form: pdf
Submission Number: 91
Loading