Vessel Analysis in Narrow Band Imaging Bronchoscopic Video
Open Access
- Author:
- Bandyopadhyay, Saptarashmi
- Graduate Program:
- Computer Science and Engineering
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- April 22, 2020
- Committee Members:
- William Evan Higgins, Thesis Advisor/Co-Advisor
Robert Collins, Committee Member
Chitaranjan Das, Program Head/Chair - Keywords:
- Narrow Band Imaging
Bronchoscopy
Vessel Enhancement
Vessel Segmentation
Artificial Intelligence
Lung Cancer Detection - Abstract:
- Lung cancer is the leading cause of cancer fatalities in the world. The 5-year survival rate of patients is only 18%, primarily because lung cancer is detected at a late stage for 75% patients, in spite of all the improvements in bronchoscopy and radiological imaging. Early detection of lung cancer can help in decreasing the mortality rate from this disease. In this thesis, we have developed methods to enhance and segment blood vessels automatically from narrow band imaging (NBI) bronchoscopic videos. The objective is to facilitate early detection of lesion growth in the airway tract in future by detecting vessels that can potentially supply blood to these lesions. The work has been motivated based on the observed limitations of subjective visual inspection of enormous volumes of data stored in NBI videos. Identification of the nonvasculature regions will be also very useful to conduct biopsy and other tests on the airway. We divide our methods into the following steps. At first, we extracted image frames from the NBI videos which are RGB images which have then been converted to HSV image frames to obtain the value channel. Then anisotropic diffusion is applied on the value channel to make the images less noisy. After that Hessian-based vessel enhancement equations have been implemented, followed by segmentation with masks, obtained from logical combinations of image clusters obtained by applying k-means clustering algorithm on the RGB image. The evaluation has been done by calculating sensitivity, specificity and accuracy performance metrics. Ground-truthing was done on some image frames manually as the initial dataset was not labeled. Our methods have achieved an aggregate statistics of 94% accuracy, 94.6% specificity and approximately 70% sensitivity while segmentation on some frames have led to very high sensitivity of 87.5% and specificity and accuracy above 96%. In this thesis, we have demonstrated that the segmented results from our vessel enhancement and segmentation techniques are quantitatively better than the existing vessel enhancement methods. This result shows the success of identifying almost all the vascular and non-vascular region which will be useful in future work on early cancer detection.