A Robust Scheme of Vertebrae Segmentation for Medical Diagnosis

Precise and reliable segmentation of vertebral column is crucial for accurate computer aided diagnostic system (CADx). It assists the radiologists and doctors in identifying various pathologies in the vertebrae with better visualizations. Accurate segmentation of vertebral disks and bones from the m...

Full description

Saved in:
Bibliographic Details
Published in:IEEE access 2019, Vol.7, p.120387-120398
Main Authors: Rehman, Faisal, Ali Shah, Syed Irtiza, Riaz, Naveed, Gilani, Syed Omer
Format: Article
Language:English
Subjects:
Artificial neural networks
Biomedical imaging
Bones
Computed tomography
computer aided diagnosis
Datasets
deep learning
Diagnostic systems
Disks
Image segmentation
Intervertebral disk
Level set
Medical imaging
Neural networks
Object segmentation
Physicians
Robustness
Shape
Task analysis
thoracolumbar vertebrae
Vertebrae
vertebrae segmentation
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Precise and reliable segmentation of vertebral column is crucial for accurate computer aided diagnostic system (CADx). It assists the radiologists and doctors in identifying various pathologies in the vertebrae with better visualizations. Accurate segmentation of vertebral disks and bones from the medical images is a tough job and becomes more challenging when dealing with various deformities and pathologies. While remarkable success was achieved by deep convolutional neural networks (DCNNs) in medical image segmentation, it is still a difficult task for DCNNs to handle the medical image segmentation problems with various deformities and anatomical complexities. In this paper, we propose a novel and efficient framework to address the subject problem by integrating a parametric level set approach in deep convolutional neural networks. The proposed scheme utilizes the probability map of pre-trained deep network to initialize the level set and it refines the output iteratively under the action of various forces to fine-tune the training of deep network. Thus the learning of the network is improved and the network is able to accommodate high topological shape variations in the vertebrae. This proposed method was evaluated on two different datasets. The first one is 20 publically available 3D spine MRI dataset to perform disc segmentation and the second one is 173 computed tomography scans for thoracolumbar (thoracic and lumbar) vertebrae segmentation. The dice score was found to be 90.37 ± 0.9 percent for disks segmentation and 94.7 ± 1.1 with ASSD of 0.1 ± 0.04 mm for thoracolumbar vertebrae segmentation. The results reveal that our proposed method is robust over multiple segmentations and outperformed the recently published state of art methods.
ISSN:2169-3536
2169-3536
DOI:10.1109/ACCESS.2019.2936492