Mentor: Dr. Joseph Reinhardt
Identifying COPD severity and predicting COPD progression from CT images can help identify the way in which an individual is affected by COPD. The structure of the airways (airway size, length, branching angle) is not commonly studied in relation to disease progression. We apply machine learning techniques to study the airways that are segmented from the CT images. Through unsupervised learning techniques such as clustering trends can emerge. In addition, I am utilizing graph neural networks to allow for more complicated comparisons to be made. In the future this will allow us to define COPD phenotypes and broadly classify how the disease progresses over time.

I use dual-energy CT to assess pulmonary parenchymal perfusion characteristics in a large multi-center study (MESA Lung) serving to phenotype risks for chronic obstructive pulmonary disease. Our working hypothesis is that individuals shutting down perfusion in inflamed parenchymal regions will be most susceptible to the development of emphysema. My role has been evaluating the images of 750 volunteers, developing metrics serving to characterize both the scan characteristics as well as to characterize the pulmonary pathologic indices. My experience with DECT of the lung has extended to evaluating similar scans of patients being evaluated for pulmonary emboli with a diagnosis of COVID-19, and utilizing our texture analysis software to evaluate non-contrast scans of COVID-19 patients for signs which seem to indicate micro-embolic disorders whereby normal-appearing lung tissues are hypo-lucent and characterized as emphysema.

Pulmonary vasculature patterns measured in high-resolution computed tomography (CT) images provide valuable features for characterizing lung diseases. The pulmonary vasculature consists of arterial and veinous subtrees which may be affected differently by different lung diseases. Artery and vein separation in CT is a challenging task due to similar intensity, close proximity, small size, and complex branching structure. The main focus of my research is to develop novel deep learning tools to automatically segment the vasculature and separate the vasculature into the arterial and veinous subtrees. These tools will be used to analyze large datasets of CT and dual-energy CT images from multi-institutional clinical trials and develop novel phenotypes of vasculature patterns in pulmonary diseases such as COPD.

Mentor: Dr. Ching-Long Lin
My research is mainly about discovering potential phenotypes of lung diseases or subgroups of patients with lung diseases, and then simulating airflow different diseased lungs. We apply digital image processing, deep learning techniques, and big data analytics on cohorts with large amount of computed tomography (CT) images to search for possible phenotypes or subgroups. With computational fluid dynamics (CFD), we are able to simulate the airflow in the lungs, hoping to provide guidance for the doctors to customize treatments for different phenotypes or subgroups and help the patients to better manage their disease. The topics that we are currently working on are Chronic Obstructive Pulmonary Disease (COPD) and Humidifier Disinfectant-associated Lung Injuries (HDLI).

Mentor: Dr. Milan Sonka
Deep learning can be a powerful tool, but the inherent “black box” nature of these methods makes it significantly more difficult to build the trust necessary to make the translation from research to clinical practice. My research focuses on developing explainable deep learning techniques for the segmentation and classification of 3 and 4-D pulmonary CT images. The generated explanations can provide the evidence and clarification necessary to help clinicians better understand and validate the decisions made by these models. As a further step, we are also researching the use of these generated explanations for anomaly detection. By detecting regions of CT images that cause inconsistent explanations, we can automatically detect abnormalities such as lung nodules or pneumothorax.