Automatic classification of the cerebral vascular bifurcations using dimensionality reduction and machine learning

This paper presents a method for the automatic labeling of vascular bifurcations along the Circle of Willis (CoW) in 3D images. Our automatic labeling process uses machine learning as well as dimensionality reduction algorithms to map selected bifurcation features to a lower dimensional space and thereafter classify them. Unlike similar studies in the literature, our main goal here is to avoid a classical registration step commonly applied before resorting to classification. In our approach, we aim to collect various geometric features of the bifurcations of interest, and thanks to dimensionality reduction, to discard the irrelevant ones before using classifiers.

In this paper, we apply the proposed method to 50 human brain vascular trees imaged via Magnetic Resonance Angiography (MRA). The constructed classifiers were evaluated using the Leave One Out Cross-Validation approach (LOOCV). The experimental results showed that the proposed method could assign correct labels to bifurcations at 96.8% with the Naive Bayes classifier. We also confirmed its functionality by presenting automatic bifurcation labels on independent images.