DeepVesselNet: Vessel Segmentation, Centerline Prediction, and Bifurcation Detection in 3-D Angiographic Volumes

We present DeepVesselNet, an architecture tailored to the challenges faced\nwhen extracting vessel networks or trees and corresponding features in 3-D\nangiographic volumes using deep learning. We discuss the problems of low\nexecution speed and high memory requirements associated with full 3-D\nconvolutional networks, high-class imbalance arising from the low percentage of\nvessel voxels, and unavailability of accurately annotated training data - and\noffer solutions as the building blocks of DeepVesselNet.\n First, we formulate 2-D orthogonal cross-hair filters which make use of 3-D\ncontext information at a reduced computational burden. Second, we introduce a\nclass balancing cross-entropy loss function with false positive rate correction\nto handle the high-class imbalance and high false positive rate problems\nassociated with existing loss functions. Finally, we generate synthetic dataset\nusing a computational angiogenesis model capable of generating vascular trees\nunder physiological constraints on local network structure and topology and use\nthese data for transfer learning.\n DeepVesselNet is optimized for segmenting and analyzing vessels, and we test\nthe performance on a range of angiographic volumes including clinical MRA data\nof the human brain, as well as X-ray tomographic microscopy scans of the rat\nbrain. Our experiments show that, by replacing 3-D filters with cross-hair\nfilters in our network, we achieve over 23% improvement in speed, lower memory\nfootprint, lower network complexity which prevents overfitting and comparable\naccuracy (with a Cox-Wilcoxon paired sample significance test p-value of 0.07\nwhen compared to full 3-D filters). Our class balancing metric is crucial for\ntraining the network and transfer learning with synthetic data is an efficient,\nrobust, and very generalizable approach leading to a network that excels in a\nvariety of angiography segmentation tasks.\n