When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.
Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.
Ischemic stroke caused by blocked arteries in the brain is one of the leading causes of death and disability worldwide. Endovascular thrombectomy treatment (EVT) is one of the best treatment strategies for restoring blood flow through blocked arteries, but its success rate depends on a number of factors, including the extent of a patient’s collateral circulation. Collateral circulation works as a radiologic surrogate predicting the response of revascularization therapy, which helps viable brain tissues to get oxygen and nutrients temporarily. Assessment of collateral circulation in ischemic stroke, which can identify patients for the most appropriate treatment strategies, is mostly conducted with visual inspection by a radiologist. Yet, numerous studies have shown that visual inspection suffers from inter- and intra-rater variability. Furthermore, the performance depends on the experience, training, and specialty of radiologists. Computer-aided decision support systems aim to provide robust methods that do not suffer from inter- and intra-rater inconsistencies. Collateral grading also suffers from the selection of appropriate cerebral imaging.
Recently, 4-dimensional computed tomography angiography (4D CTA) has become a reliable medium of cerebral vasculature imaging to give details of collateral circulation. It helps to avoid the inaccurate estimation of collaterals, unlike single-phase CTA, which is prone to sub-optimal selection of a time point for scanning. Along with 4D CTA, non-contrast computed tomography (NCCT) is easily available and used as a front-line diagnostic tool in clinical settings, being free from contrast agents that can cause adverse effects to some patients. Therefore, we propose computer-aided automatic systems for collateral evaluation in ischemic stroke using 4D CTA and NCCT imaging. Herein, we propose an automatic quantification method considering the minimization problem, classic machine learning (ML) as well and deep learning (DL) methods for the automatic evaluation of collaterals. Unlike traditional ML methods, DL models can automatically learn and extract relevant features from data, reducing the need for manual feature engineering. Nonetheless, DL methods necessitate substantial data, which is scarce in the context of ischemic stroke. Consequently, the task of collateral evaluation encounters the dual challenge of data scarcity and class imbalance. To address this, we have developed DL models utilizing transfer learning, where we manage class imbalance by employing focal loss with class weights to penalize the majority class. However, the performance of transfer learning networks is hampered by the limited availability of pre-trained networks in the medical domain. To overcome this limitation, we have explored the potential of few-shot classification and Siamese networks. These innovative approaches capitalize on their ability to effectively generalize from a limited labeled dataset. Therefore, we have considered these methods to develop robust approaches that facilitate automatic collateral evaluation while mitigating the challenges posed by a small, imbalanced dataset.