Please use this identifier to cite or link to this item:
Title: Predicting Clinical Outcome in Acute Ischemic Stroke Using Parallel Multi- Parametric Feature Embedded Siamese Network
Authors: Osama, Saira
Keywords: Computer Science
Computer & IT
Issue Date: 2021
Publisher: National University of Computer and Emerging Sciences Islamabad
Abstract: Stroke is the secondary reason for global disability and death. Of all stroke types, ischemic stroke occurs the most, due to restricted blood flow in the brain. The ischemic stroke passes through different disease stages. The early stage is the acute stage. The usual practice of diagnosing acute stroke involves Multi-parametric Magnetic Resonance Imaging (mpMRI) within a limited time window. The treatment at the acute stage must be administered within this time window and involves high risk. mpMRI can detect and locate the stroke lesion at the acute stage. Besides, it presents the blood flow situation within the brain, helping in measuring the gains and risks involved in treatments. The benefits and risks can be accessed via clinical outcome measured at 3- month follow-up in the form of a modified Rankin Scale (mRS). Assessment of outcomes of risky therapies using mpMRI beforehand i.e., before any treatment can be of great value in clinical practice. However, it is quite complex because of varying stroke lesion characteristics. These characteristics include location, size, shape, and cerebral hemodynamics involved. It is also challenging as mpMRI involves multiple MRI scans of a patient to be analyzed at the same time thus making the analysis suffer from inter-variability and intra-variability. A fully automated machine learning model that can predict clinical outcomes using mpMRI at an initial stage can provide invaluable evidence for treatment decisions. However, the mpMRI datasets for training the models especially deep learning models are rare and have a high-class imbalance. Moreover, the evaluation metrics used in measuring the performances of the state-of-the-art methods in previous studies are sensitive to the imbalance and might mislead in the assessment of models’ performances. This dissertation presents a novel deep learning model named Parallel Multiparametric Feature Embedded Siamese Network (PMFE-SN). The model can learn with very limited training examples along with handling imbalance in multi-parametric MRI data. Furthermore, the performance of PMFE-SN is evaluated in contrast to previous state of-the-art methods, using metrics that are not sensitive to imbalance. The results obtained using all the metrics demonstrate that PMFE-SN has surpassed other state-of-the-art methods. The developed model predicts the minority as well as majority class with almost equal accuracy. All results are computed using leave one cross out testing. PMFE-SN attained an accuracy of 0.67 for the class with a minimum number of data points (minority class), only two used to train the model. 0.61 is the accuracy obtained for the majority class, with the highest count of examples. Comparing to PMFE-SN, the top-ranked methods utilizing hand-crafted features (radiomics) achieved 0.33 and 0 accuracy for the majority and minority class, respectively.
Gov't Doc #: 25934
Appears in Collections:PhD Thesis of All Public / Private Sector Universities / DAIs.

Files in This Item:
File Description SizeFormat 
saira osama CS 2021 nu fast isb.pdfphd.Thesis3.97 MBAdobe PDFView/Open

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.