JavaScript is disabled for your browser. Some features of this site may not work without it.
Automated analysis of quantitative image biomarkers from low-dose chest CT scans

Author
Liu, Shuang
Abstract
A quantitative imaging biomarker is a quantitatively measured characteristic derived from medical images, which serves as cost¬-effective and non¬invasive tools for patient health assessment, including diagnosis and periodic screening of disease, therapy planning as well as longitudinal monitoring of treatment response. This dissertation presents an automated framework for quantitative image biomarker measurement and evaluation from the low-dose chest CT (LDCT) scans that are acquired during the annual lung cancer screening. Four categories of quantitative image biomarkers are investigated, including breast density and gynecomastia quantification, bone mineral density (BMD), airway dimensions and pulmonary nodule classification. An anatomy directed approach is applied to the analysis of the breast region and to the biomarker measurements. The fully automated breast density assessment and gynecomastia measurements have been demonstrated to be consistent with the reading of radiologists. Fully automated BMD assessment is achieved by building upon the model-based segmentation and anatomical labeling of individual vertebral body. Statistically significant strong correlation with the gold standard reference can be obtained at all vertebral levels. A fully automated knowledge-based approach is applied to the segmentation and anatomical labeling of each airway bronchus, which enables the measurements of precise and reproducible airway dimensions. For the classification of pulmonary nodule malignancy, a 3D CNN is trained from scratch and demonstrates various advantages over both the traditional machine learning approaches using hand-crafted 3D image features and the 2D CNN models. Classifier ensembles of the combinations of the 3D CNN and traditional machine learning models achieve the best performance by taking advantage of the complementary characteristics of the traditional models and the CNN models. In conclusion, with the recent large-scale implementation of annual lung cancer screening in the US using LDCT, great potential emerges for the concurrent extraction of quantitative image biomarkers from different regions in the chest, which are covered in LDCT. This dissertation has demonstrated the feasibility of fully automated measurement and evaluation of a rich set of quantitative image biomarkers, and the opportunity to significantly enhance the impact of LDCT by offering a comprehensive health assessment to each screening participant with no additional imaging or radiation exposure.
Date Issued
2018-08-30Subject
machine learning; Computer engineering; Deep Learning; Electrical engineering; Computer science; automated medical imaging analysis; computer-aided diagnosis; low-dose chest CT; lung cancer screening
Committee Chair
Reeves, Anthony P.
Committee Member
Doerschuk, Peter; Chen, Tsuhan
Degree Discipline
Electrical and Computer Engineering
Degree Name
Ph. D., Electrical and Computer Engineering
Degree Level
Doctor of Philosophy
Type
dissertation or thesis