Pulmonary emphysema contributes to the chronic airflow limitation characteristic of chronic obstructive pulmonary disease (COPD), which is a leading cause of morbidity and mortality worldwide. Computed tomography (CT) has enabled in vivo assessment of pulmonary emphysema at the macroscopic level, and is commonly used to identify and assess the extent of the disease.
During the past decade, the availability of CT imaging data has increased rapidly, while the image quality has continued to improve. High-resolution CT is extremely valuable both for patient diagnosis and for studying diseases at the population level. However, visual assessment of these large data sets is subjective, inefficient, and expensive. This has increased the demand for objective, automatic, and reproducible image analysis methods.
For the assessment of pulmonary emphysema on CT, computational models usually aim either to give a measure of the extent of the disease, or to categorize the emphysema subtypes apparent in a scan. The standard methods for quantitating emphysema extent are widely used, but they remain sensitive to changes in imaging protocols and patient inspiration level. For computational subtyping of emphysema, the methods remain at a developmental stage, and one of the main challenges is the lack of reliable label data. Furthermore, the classic emphysema subtypes were defined on autopsy before the availability of CT and could be considered outdated. There is also no consensus on how to match the subtypes on autopsy to the varying emphysema patterns present on CT.
This work presents two methodological improvements for analyzing emphysema on CT. For the assessment of emphysema extent, a novel probabilistic approach is introduced and evaluated on a longitudinal data set with varying imaging protocols. The presented model is shown to improve significantly compared to standard methods, particularly at the presence of differing noise levels. The approach is also applied on quantifying emphysema on a large data set of cardiac CT scans, and is shown to improve the prediction of emphysema extent on subsequent full-lung CT scans.
The second major contribution of this work applies unsupervised learning to recognizing patterns of emphysema on CT. Instead of trying to reproduce the classic subtypes, the novel approach aims to capture the most dominant variations of lung structure pertaining to emphysema. While removing the reliance on visually assigned labels, the learned patterns are shown to represent different manifestations of emphysema with distinct appearances and regular spatial distributions. The clinical significance of the patterns is also demonstrated, along with high-level performance in the application of content-based image retrieval.
The contributions of this work advance the analysis of emphysema on CT by applying novel machine learning approaches to increase the value of the available imaging data. Probabilistic methods improve from the crude standard methods that are currently used to quantitate emphysema, and the value of learning disease patterns directly from image data is demonstrated. The common framework relying on replicating visually assigned labels of outdated subtypes has not achieved widespread acceptance. The methodology presented in this work may have a substantial impact on how emphysema subtypes on CT are recognized and defined in the future.
| Identifer | oai:union.ndltd.org:columbia.edu/oai:academiccommons.columbia.edu:10.7916/D8JW8CPH |
| Date | January 2015 |
| Creators | Häme, Yrjö |
| Source Sets | Columbia University |
| Language | English |
| Detected Language | English |
| Type | Theses |
Page generated in 0.0016 seconds