Deep learning-based semantic vessel graph extraction for intracranial aneurysm rupture risk management

Abstract

Purpose Intracranial aneurysms are vascular deformations in the brain which are complicated to treat. In clinical routines, the risk assessment of intracranial aneurysm rupture is simplified and might be unreliable, especially for patients with multiple aneurysms. Clinical research proposed more advanced analysis of intracranial aneurysm, but requires many complex preprocessing steps. Advanced tools for automatic aneurysm analysis are needed to transfer current research into clinical routine. Methods We propose a pipeline for intracranial aneurysm analysis using deep learning-based mesh segmentation, automatic centerline and outlet detection and automatic generation of a semantic vessel graph. We use the semantic vessel graph for morphological analysis and an automatic rupture state classification. Results The deep learning-based mesh segmentation can be successfully applied to aneurysm surface meshes. With the subsequent semantic graph extraction, additional morphological parameters can be extracted that take the whole vascular domain into account. The vessels near ruptured aneurysms had a slightly higher average torsion and curvature compared to vessels near unruptured aneurysms. The 3D surface models can be further employed for rupture state classification which achieves an accuracy of 83.3%. Conclusion The presented pipeline addresses several aspects of current research and can be used for aneurysm analysis with minimal user effort. The semantic graph representation with automatic separation of the aneurysm from the parent vessel is advantageous for morphological and hemodynamical parameter extraction and has great potential for deep learning-based rupture state classification.

Classification, detection, and segmentation performance of image-based AI in intracranial aneurysm: a systematic review

“…In addition, 16 studies [ 22 – 29 , 4 , 30 , 9 , 31 – 35 ] utilized a pattern classification based on image features, aiming to extract deep semantic features from images using deep learning models and utilize these features for automatic classification and recognition of IAs. The test group’s average accuracy ranged from 74.5% to 98.8%, with average sensitivity ranging from 48.3% to 99.3% and average specificity ranging from 18.2% to 98.1%.…”