Major
Data Science
Research Abstract
Neuro-radiologists currently use qualitative volumetric change of brain ventricles after surgery to assess the safety of removing a ventriculoperitoneal (VP) shunt which is a medical device that relieves pressure on the brain caused by fluid accumulation. Following safe removal of the VP shunt, patients can be released from the hospital. The need for accurate and quick measurement of brain ventricular volumetric change makes automatic 3D segmentation software an ideal candidate to aid decisions after surgery. In this paper, we propose an approach to estimate the ventricular volume variation using segmentation in brain MRI and CT images. Our approach consists of using cascaded models, each of which are based on 3D U-Net, a Convolutional Neural Network (CNN) based architecture. In the first step we input a preprocessed 3D image in the model to obtain as output the binary volume segmented brain mask. This mask is used for skull stripping the original image to remove excess noise to improve ventricle segmentation performance. The skull stripped image is the input in the next 3D U-net to obtain the ventricles mask. This two-step approach allow us to estimate the ventricles volume variation value between two consecutive patient images, computing each ventricles volume value from the respective masks. The same approach was replicated independently for MRI and CT scans. Our two-step deep learning segmentation approach was evaluated against 2 test sets for CT images and 2 test sets for MR images that varied in segmentation difficulty. In addition, we compared performance to widely-used atlas-based segmentation and found our deep learning approach is capable of segmenting ventricles at higher dice scores in both the MR and CT test cases.
Faculty Mentor/Advisor
Yannet Interian
Included in
Automated Segmentation of Brain Ventricles Using 3D U-NET
Neuro-radiologists currently use qualitative volumetric change of brain ventricles after surgery to assess the safety of removing a ventriculoperitoneal (VP) shunt which is a medical device that relieves pressure on the brain caused by fluid accumulation. Following safe removal of the VP shunt, patients can be released from the hospital. The need for accurate and quick measurement of brain ventricular volumetric change makes automatic 3D segmentation software an ideal candidate to aid decisions after surgery. In this paper, we propose an approach to estimate the ventricular volume variation using segmentation in brain MRI and CT images. Our approach consists of using cascaded models, each of which are based on 3D U-Net, a Convolutional Neural Network (CNN) based architecture. In the first step we input a preprocessed 3D image in the model to obtain as output the binary volume segmented brain mask. This mask is used for skull stripping the original image to remove excess noise to improve ventricle segmentation performance. The skull stripped image is the input in the next 3D U-net to obtain the ventricles mask. This two-step approach allow us to estimate the ventricles volume variation value between two consecutive patient images, computing each ventricles volume value from the respective masks. The same approach was replicated independently for MRI and CT scans. Our two-step deep learning segmentation approach was evaluated against 2 test sets for CT images and 2 test sets for MR images that varied in segmentation difficulty. In addition, we compared performance to widely-used atlas-based segmentation and found our deep learning approach is capable of segmenting ventricles at higher dice scores in both the MR and CT test cases.
