Year: 2026
Type: dissertation
Abstract: Intracranial aneurysms (IAs) are dilations in intracranial blood vessels, affecting approximately 3% of the population, with unruptured aneurysms increasingly found in unrelated images due to the increase in medical imaging in general. IAs are commonly found using computed tomography angiography (CTA). Approximately 20% to 50% of IAs rupture, causing subarachnoid hemorrhage, which is a life-threatening condition. Unruptured aneurysms can be treated, but treatments come with substantial risks. Thus, it is of paramount importance to determine the risks of rupture before any treatment. Consequently, the increase in medical imaging of the head calls for an automatic tool to find and segment IAs from these images. Deep learning based methods can be used to detect and segment IAs for diagnosis and rupture risk stratification purposes. This thesis uses two segmentation architectures, U-Net and Swin-UNETR, to detect and segment IAs from CTA images using a coarse-to-fine approach. The coarse-to-fine approach was chosen to mitigate the large class imbalance between the small aneurysm and the large volume of the rest of the image. To improve the clinical reliability and usability of the models, Monte-Carlo dropout and deep ensembling, are applied to quantify the predictive uncertainty of the models. The evaluation results of the coarse model deep ensembles showed a sensitivity of 0.881 to 0.940, the model correctly detects the vast majority of aneurysms. The Dice score was 0.811 to 0.818 for the segmentations, produced by the fine model. The results were comparable to those of the state-of-the-art models evaluated on the same dataset. Crucially, the uncertainty quantification results indicate that the model’s performance is weakly correlated with its uncertainty. The model was most certain about cases with the best quality segmentation. These results reflect the scalability and robustness of existing segmentation architectures. Instead of creating a highly application-specific architecture or trying to build in task related information in the architecture, deep ensembles of large versions of more universal segmentation models were trained. This approach created a simple but effective pipeline. Future work on this topic will focus on training larger ensembles with more data, evaluating with other uncertainty metrics, as well as more comprehensive comparisons with other models.
Source: Aaltodoc (Aalto University)
Author: Ruuskanen, Linnea
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Funder: Business Finland
OpenAlex ID: W7134760897