Deep Learning for Perfusion Cerebral Blood Flow (CBF) and Volume (CBV) Predictions and Diagnostics

Abstract

AbstractDynamic susceptibility contrast magnetic resonance perfusion (DSC-MRP) is a non-invasive imaging technique for hemodynamic measurements. Various perfusion parameters, such as cerebral blood volume (CBV) and cerebral blood flow (CBF), can be derived from DSC-MRP, hence this non-invasive imaging protocol is widely used clinically for the diagnosis and assessment of intracranial pathologies, including tumor classification, stroke assessment, seizure detection, etc. Currently, most institutions use commercially available software to compute the perfusion parametric maps. Conventionally, the parametric maps are derived by mathematical equations which require the selection of vascular input waveforms. However, these conventional methods often have limitations, such as being time-consuming and sensitive to user input, which can lead to inconsistent results; this highlights the need for a more robust and efficient approach like deep learning. Using relative cerebral blood volume (rCBV) and relative cerebral blood flow (rCBF) perfusion maps generated by an FDA-approved commercial software, we trained a multi-step deep learning (DL) model. The model used each 4D MRP dataset as input, and combined the temporal features extracted from each voxel with spatial information of the brain to predict voxel-wise perfusion parameters. DL-generated rCBV and rCBF maps were evaluated qualitatively and quantitatively. An auxiliary (control) model, with similar architecture, but trained with truncated datasets that had fewer time points, was designed to explore the contribution of temporal features. Our model is based on a multistage encoder-decoder architecture that leverages a 1D convolutional neural network (CNN) as the first encoder to capture temporal information, followed by a 2D U-Net encoder-decoder network to process spatial features. This combination of encoders allows our model to effectively integrate time-varying and spatial data, generating accurate and comprehensive CBV/CBF predictions for the entire brain volume. Our model demonstrates comparable results to that of FDA-approved commercial software.