Author ORCID Identifier
0000-0002-2748-1444
Date of Graduation
8-2017
Document Type
Thesis (MS)
Program Affiliation
Medical Physics
Degree Name
Masters of Science (MS)
Advisor/Committee Chair
Laurence Court
Committee Member
Ann Klopp
Committee Member
Erik Cressman
Committee Member
Stephen Kry
Committee Member
Jennifer Johnson
Committee Member
Arvind Rao
Committee Member
Jinzhong Yang
Abstract
The mortality rate of cervical cancer is approximately 266,000 people each year, and 70% of the burden occurs in Low- and Middle- Income Countries (LMICs). Radiation therapy is the primary modality for treatment of locally advanced cervical cancer cases. In the absence of high quality diagnostic imaging needed to identify nodal metastasis, many LMIC sites treat standard pelvic fields, failing to include node metastasis outside of the field and/or to boost lymph nodes in the abdomen and pelvis. The first goal of this project was to create a program which automatically identifies positive cervical cancer lymph nodes on non-contrast daily CT images, which are widely available in LMICs(1).
A region of interest which is likely to contain the nodal volumes relevant for cervical cancer was defined on a single patient CT(2). This region was deformed onto new patients using an in-house, demons-based deformation software. Edge detection and erosion filtering were used to distinguish potential positive nodes from normal structures. Regions on adjacent slices were then connected into a potential nodal 3D-structure. To differentiate these 3D structures from normal tissues, eighty-six features were generated based on the shape and mean pixel values of the structures, and four classification ensemble methods were tested to differentiate the positive nodes from normal tissues. A cohort of fifty-eight MD Anderson cervical cancer patients with pathologically enlarged lymph nodes were used as a training-test set. Similarly, twenty MD Anderson cervical cancer patients were obtained as a validation set. They contained 154 and 35 pathologically enlarged lymph nodes, respectively.
Model comparison led to the selection of the Adaboost ensemble model, utilizing 17 features. In the validation set, 60% of the clinically significant positive cervical cancer nodes were identified along with a false/true positive ratio of ~4:1. The entire process takes approximately 10/number-of-cores-minutes.
Our findings demonstrated that our computer-aided detection model can assist in the identification of metastatic nodal disease where high quality diagnostic imaging is not readily available. By identifying these nodes, radiation treatment fields can be modified to include pathologically enlarged lymph nodes, which is an essential element to providing potentially curative radiotherapy for cervical cancer.
Keywords
Cervical Cancer, CAD, Machine Learning