Abstract WP60: Kernel Spectral Regression and Neural Networks Enable Regional Detection of Hemorrhagic Transformation on Multi-Modal MRI for Acute Ischemic Stroke

Abstract

Purpose: Hyperperfusion detected on arterial spin labeling (ASL) cerebral blood flow (CBF) images acquired after acute ischemic stroke (AIS) onset has been shown to be significantly associated with development of intracerebral hemorrhage. ASL data is often too noisy for accurate voxel-based analysis alone. The purpose of this study was to develop a machine learning model for the voxel-based detection of tissue at risk for hemorrhagic transformation (HT) utilizing multiple MRI modalities. Materials and Methods: The present study utilized clinical MRI and ASL image data acquired from 67 AIS patients shortly after endovascular therapy. A novel regional cuboid sampling framework was developed for machine learning training, in which local cuboids were extracted from the CBF map, DWI, FLAIR, and TSE images before being matched with GRE-based manually drawn bleed groundtruth delineations. Kernel spectral regression (KSR) uses the eigenvectors of the graph representation to reveal low dimensional structure in high dimensional data. A two-layer feed-forward neural network was then built with 10 neurons in the sigmoid hidden layer and trained with scaled conjugate gradient backpropagation to classify cuboid inputs into likelihood of HT. Results: The proposed multimodal regional framework reached an accuracy of 80.59 ± 3% in detecting hemorrhage with KSR on our dataset (better than a voxel-based prediction on CBF with 72.80 ± 5% accuracy). Using the neural network training, the framework reached an improved accuracy of 95.1% ± 0.6%. Figure 1 shows how the regions with high likelihood of hemorrhage match well with the manually drawn regions in the reference GRE map. Conclusion: Kernel spectral regression and neural networks can provide more accurate detection of tissues at risk for HT. Although CBF can inform AIS patient clinical outcome, the addition of multi-modal MRI data into the regional cuboid framework substantially improves the voxel-based HT detection accuracy.