Abstract:
Stroke is a leading cause of death and long-term disability worldwide, which fosters
the need for effective methods of diagnosis with lesion segmentation and prediction of
recovery. Hence, this thesis presents a comprehensive approach that integrates deep
learning techniques for stroke lesion segmentation from functional magnetic resonance
imaging (fMRI) data, coupled with machine learning methods such as Random Forest
(RF) and Support Vector Machine (SVM) for predicting recovery outcomes post-stroke.
The first phase of the thesis focuses on developing a reliable stroke lesion segmentation model comparable to the state-of-the-art deep learning architectures using a large
public dataset. Specifically, families of Unets, derived from convolutional neural networks (CNNs) are employed to accurately and reliably delineate stroke lesions. These
range from transfer learning of existing theoretical concepts such as Linear kernel attention (LKA), building model using a state-of-the-art method like transformers and dense
blocks to attempting sophisticated methods such as diffusion. The adoption of enhanced
ensemble method also contributed to generating a high performing segmented label with
significantly high level of accuracy. Hence, our final segmentation model achieved an
accuracy of 0.6513 as well as sufficient reliability in delineating stroke lesions and was
comparable to the state-of-the-art method.
In the second phase, machine learning models, including random forest (RF) and Support
vector machines (SVM), are employed to predict recovery outcomes following stroke based
on extracted imaging features and clinical data from the Functional MRI data from two
different sites. The delineated lesion using the segmentation model developed in the first
phase as well as additional clinical variables to forecast functional recovery were used to
predict the recovery of the patients, and highlighted the importance of lesion volume and
its position relative to the somatosensory region of the brain for achieving high accuracy.
Overall, this thesis contributes to the advancement of stroke management by providing
reliable tools for lesion segmentation and recovery prediction. The integrated approach
combining deep learning with machine learning techniques offers a comprehensive framework for understanding stroke pathology and guiding clinical decision-making, ultimately
improving patient care and outcomes in the post-stroke rehabilitation process.