Cerebrovascular diseases are second most common cause of death, therefore their early diagnosis and treatment are crucial for preventing adverse health events or even death. In this work, we focused on anatomical labeling of cerebral arteries that enables their cross-sectional quantification and inter-subject comparisons, e.g. healthy versus pathological. Such comparisons are the basis to identify geometric risk factors correlated to the cerebrovascular diseases. Automation of anatomical labeling is an important step towards higher accuracy, reliability and efficiency of the labeling process and is needed for its practical implementation.
We (re)implemented and evaluated four 3D deep learning methods to label the main arteries comprising the Circle of Willis, e.g. the region with the highest incidence of cerebral aneurysms. The four tested methods were PointNet, PointNet++, DGCNN and HodgeNet. In comparison with existing Circle of Willis labeling methods, the tested deep learning methods were very efficient as they did not require image preprocessing in form of vessel extraction, a process than generally involves manual corrections. For test and evaluation purposes, we used 165 intracranial magnetic resonance angiography (MRA) images, in which we manually labeled the Circle of Willis arteries. In addition we tested and evaluated the impact of additional features, for example the MRA signal intensity, on the results of labeling. The 3D deep learning methods achieved accuracy between 85 and 95 \% in terms of Dice overlap (PointNet++), which was comparable to the results of most existing methods, while the labeling process took only couple of seconds on average. The use of additional features improved the labeling results for 3\%. The results show great potential of the 3D deep learning methods for carrying out the anatomical labeling task in a practical setting.
|
