Image Analysis and Deep Learning Techniques for the Detection and Characterisation of Unruptured Intracranial Aneurysms

Abstract

This thesis presents image analysis techniques for the detection and characterisation of unruptured intracranial aneurysms (UIAs). Using such methods, robust and reliable growth and rupture assessment of UIAs can be made to aid in treatment decision making. The first half of this thesis consider automatic UIA detection and segmentation methods from angiographic scans (TOF-MRAs and CTAs). Chapter 2 describes the organisation of an international biomedical image analysis challenge. Teams submitted automatic UIA detection and segmentation methods, which were evaluated on a held-out test set. The winning detection method has since been developed into an open-source framework for medical image detection (nnDetection). The challenge remains as an important benchmark for UIA detection and segmentation methods. Chapter 3 describes an anomaly detection method using a variational autoencoder (VAE) trained on healthy TOF-MRAs. Reconstructed TOF-MRAs with diagnosed aneurysms had a lower Structural Similarity Index (SSIM), than TOF-MRAs of subjects with no aneurysms. Importantly, the results identified that structure and shape within the scans, and not just intensity, is important for UIA detection. The UIA detection method in Chapter 4 exploits the fact surface of a UIA is different from surrounding vessels. Vessels were segmented from TOF-MRAs and meshes were fitted to the surface. A mesh convolutional neural network was trained using the labelled vessel meshes, to detect UIAs. The modality-independent method has comparable performance for both TOF-MRAs and CTAs. Automatic UIA detection and segmentation described in Chapters 2-4 allow automatic 3D volume and morphology UIA measurements for potential growth and rupture risk assessment. Chapters 5 and 6 investigate UIA volume and morphology and their relationship to UIA growth assessment. In Chapter 5, 3D volume growth assessment was more reliable than 2D size, with smaller interobsever differences, and more consistency across location. The smallest detectable change for 2D growth was larger than the current growth definition (1mm), leading to ambiguity in the current definition. 3D UIA quantitative morphology measures, such as flatness and shape index, were introduced in Chapter 6 and their relationship with UIA growth investigated. Even in non-growing (stable) aneurysms, morphology changed, suggesting that non-growing aneurysms could still be unstable. Quantified morphologic change should be considered for UIA growth and rupture risk assessment. Finally, in Chapter 7, a UIA growth prediction model using a vessel surface mesh convolutional neural network was developed. The model had comparable performance to patient demographic growth prediction models (ELAPSS). Mesh/morphology and patient models could be combined to provide a complete UIA growth prediction model. In conclusion, this thesis provides complete UIA characterisation from TOF-MRAs using computer-aided techniques. The automatic UIA detection and segmentation allows reliable, automatic UIA volume and morphology measurements. Such measurements aid in UIA growth assessment and formal UIA growth definitions including these measures should be investigated. As the accuracy of automatic UIA segmentation methods and growth prediction models increase, these will become more commonplace in clinical workflows. This could result in a fully automatic UIA characterisation tool, determining UIA volume, morphology and growth prediction scores to aid in treatment decision and improve patient outcome.