Graduated: January 1, 2014
Feature Engineering for 3D Medical Image Applications
eature engineering, including input representation, feature design, evaluation, and optimization, is essential to success in machine learning. For unstructured data like images and texts, feature engineering can often become the bottleneck in learning related tasks. Selecting the most effective and descriptive features can improve performance, proficiency, and precision in quantification applications, or enhance a good classifier in classification. Features are domain-specific, therefore, deciding what features to use and optimizing the design so that features can express input explicitly, automatically, fully, yet intuitively often require substantial knowledge of the applications and the nature of the input. This thesis introduces a new set of feature engineering algorithms for medical research of 3D CT skull images in understanding craniosynostosis disorder. Three related tasks: 1) classification, 2) severity assessment and class ranking, and 3) pre-post surgery change are used to demonstrate the effectiveness of the features and the algorithms that produce them.
Craniosynostosis, a disorder in which one or more fibrous joints of the skull fuse prematurely, causes skull deformity and is associated with increased intracranial pressure and developmental delays. In order to perform medical research studies that relate phenotypic abnormalities to outcomes such as cognitive ability or results of surgery, biomedical researchers need an automated methodology for quantifying the degree of abnormality of the disorder. While several papers have attempted this quantification through statistical models, the methods have not been intuitive to biomedical researchers and clinicians who want to use them. The goal of this work was to develop a general set of features upon which new quantification measures could be developed and tested. The features reported in this study were developed as basic shape measures, both single-valued and vector-valued, that are extracted from a projection-based plane of the 3D skull. This technique allows us to process images that would otherwise be eliminated in previous systems due to poor resolution, noise or imperfections on their original older CT scans.
We test our new features on classification tasks and also compare their performance to previous research. In spite of their simplicity, the classification accuracy of our new features is significantly higher than previous results on head CT scan data from the same research studies.
We propose a set of features derived from CT scans of the skull that can be used to quantify the degree of abnormality of the disorder. A thorough set of experiments is used to evaluate the features as compared to two human craniofacial experts in a ranking evaluation.
We study pre-post surgery change based on selected features we use in quantifying the severity of deformity of the disorder. Using the same selected features, we also compare and contrast post-surgery craniosynostosis skulls to the unaffected class.
Linda G. Shapiro (Chair), Michael L. Cunningham, Su-In Lee, Helen Sherk (GSR)