Stroke Lesion Outcome Prediction Based on 4D CT Perfusion Data Using Temporal Convolutional NetworksDownload PDF

Published: 31 Mar 2021, Last Modified: 16 May 2023MIDL 2021Readers: Everyone
Keywords: stroke, outcome prediction, CT perfusion, deep learning, temporal convolutional network
TL;DR: We developed and evaluated a temporal convolutional neural network to predict stroke lesion outcomes directly from 4D CTP datasets without computing perfusion maps.
Abstract: Acute ischemic stroke is caused by a blockage in the cerebral arteries, resulting in long-term disability and sometimes death. To determine the optimal treatment strategy, a patient-specific assessment is often based on advanced neuroimaging data, such as spatio-temporal (4D) CT Perfusion (CTP) imaging. To date, perfusion maps are typically calculated from 4D CTP data and then thresholded to localize and quantify the stroke lesion core and tissue-at-risk. A few studies have recently developed deep learning methods to predict stroke lesion outcomes from perfusion maps. The basic idea of these is to train a model, using perfusion maps acquired at baseline and their corresponding follow-up images acquired several days after treatment, to automatically estimate the final lesion location and volume in new patients. Nevertheless, model training based on the original 4D CTP scans might be desirable, as they could contain more valuable information not directly represented in perfusion maps. Therefore, we aimed to develop and evaluate a temporal convolutional neural network (TCN) to predict stroke lesion outcomes directly from 4D CTP datasets acquired at admission, without computing any perfusion maps. Using a total of 176 CTP scans, we investigated the impact of the time window size by training the proposed TCN on various numbers of CTP frames: 8, 16, and 32 time points. For comparison purposes, we also trained a convolutional neural network based on perfusion maps. The results show that the model trained on 32 time points yielded significantly higher Dice values (0.33$\pm$0.21) than the models trained on 8 time points (0.25$\pm$0.20; P$<$0.05), 16 time points (0.28$\pm$0.21; P$<$0.001), and perfusion maps (0.23$\pm$0.18; P$<$0.05). These experiments demonstrate that the proposed model effectively extracts spatio-temporal data from CTP scans to predict stroke lesion outcomes, which leads to better results than using perfusion maps.
Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.
Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.
Paper Type: both
Primary Subject Area: Segmentation
Secondary Subject Area: Detection and Diagnosis
Source Latex: zip
14 Replies