The design and implementation of deep learning methods for the diagnosis and anatomical analysis of stroke in brain CTs

Lieberman, Isabel

The design and implementation of deep learning methods for the diagnosis and anatomical analysis of stroke in brain CTs

Thesis / Dissertation

2026

Publisher

University of Cape Town

Department

Division of Clinical Anatomy and Biological Anthropology

Faculty

Faculty of Health Sciences

Abstract

Introduction Stroke is a leading cause of death globally, ranking second worldwide and third in South Africa. The burden on sub-Saharan African health systems is exacerbated by limited access to medical care, particularly radiological services. The application of artificial intelligence (AI) to medical diagnostics presents an exciting opportunity to address these challenges. This study aims to expand knowledge and infrastructure development for AI in medical diagnostics. Methods A dataset of publicly-available head computed tomography (CT) scans, including haemorrhagic and ischaemic stroke cases and controls, was used. Initially, SynthSeg, a deep learning model for automated segmentation and parcellation of Magnetic Resonance Imaging (MRI), was applied to estimate regional volumes, complemented by manual evaluation of segmentations. Regional volumes were compared between stroke types and to controls. Following this, two deep-learning models were developed: a modified OzNet for stroke classification and a modified Deeplabv3 for segmenting ischaemia and haematomas. Results Haemorrhagic stroke (HS) often involved multiple vascular territories (particularly middle cerebral artery/ posterior cerebral artery (MCA/ PCA)), whereas ischaemic stroke (IS) generally caused localized MCA infarcts. IS volumes were significantly larger in MCA-supplied regions, while HS volumes resembled controls except in the precentral gyrus, pallidum, and pericalcarine cortex. Both stroke types showed reduced CSF but enlarged temporal horn and third ventricle volumes. SynthSeg's performance declined for severe lesions, highlighting the challenges in volumetric analysis for advanced pathologies. The classification model achieved 0.49 AUC (indicating overfitting), while the segmentation model attained a 0.4 Dice Score, underscoring the need for refined data and further honing. Conclusion These findings highlight distinct stroke-related volume changes and segmentation challenges, emphasizing the need for pathology-adapted solutions. Enhanced datasets and computational resources could further improve AI-driven stroke diagnostics, offering a scalable option in settings with limited radiological expertise.

Keywords

brain CT's

analysis

Reference:

Collections

Masters

Full item page