Using 3-D CNNs and Local Blood Flow Information to Segment Cerebral Vasculature
Source of Publication
2018 IEEE International Symposium on Signal Processing and Information Technology, ISSPIT 2018
© 2018 IEEE. The variability of the strength (increase or decrease) of the blood flow signals throughout the range of slices of the MRA volume is a big challenge for any segmentation approach because it introduces more inhomogenities to the MRA data and hence less accuracy. In this paper, a novel cerebral blood vessel segmentation framework using Time-Of-Flight Magnetic Resonance Angiography (TOF-MRA) is proposed to handle this challenge. The segmentation framework is based on using three dimensional convolutional neural networks (3D-CNN) to segment the cerebral blood vessels taking into account the variability of blood flow signals throughout the MRA scans. It consists of the following two steps: i) bias field correction to handle intensity inhomogeneity which are caused by magnetic settings, ii)instead of constructing one CNN model for the whole TOF-MRA brain, the TOF-MRA volume is divided into two compartments, above Circle of Willis (CoW) and at and below CoW to account for blood flow signals variability across the MRA volume's slices, then feed these two volumes into the three dimensional convolutional neural networks (3D-CNN). The final segmentation result is the combination of the output of each model. The proposed framework is tested on in-vivo data (30 TOF-MRA data sets). Both qualitative and quantitative validation with respect to ground truth (delineated by an MRA expert) are provided. The proposed approach achieved a high segmentation accuracy with 84.37% Dice similarity coefficient, sensitivity of 86.14%, and specificity of 99.00%.
Institute of Electrical and Electronics Engineers Inc.
Cerebrovascular, CNN, Deep Learning, MRA
Kandil, Heba; Soliman, Ahmed; Taher, Fatma; Mahmoud, Ali; Elmaghraby, Adel; and El-Baz, Ayman, "Using 3-D CNNs and Local Blood Flow Information to Segment Cerebral Vasculature" (2019). All Works. 3847.
Indexed in Scopus