Projects - Summer 2019
Predicting Age-Related Macular Degeneration Disease from SD-OCT Images
Radiomics-Based Texture Analysis of Idiopathic Pulmonary Fibrosis
Applying Topic Modeling Techniques to Detect Trends in Mental Discussion
Predicting Age-Related Macular Degeneration Disease from SD-OCT Images
Advanced form of age-related macular degeneration (AMD) is a major health burden that can lead to irreversible vision loss in the elderly population. For early preventative interventions, there is a lack of effective tools to predict the prognosis outcome of advanced AMD because of the similar visual appearance of retinal image scans in the early stage and the variability of prognosis paths among patients. An early characteristic of AMD is drusen, which appears as yellowish deposits under the retina. AMD is mainly categorized into two types: Dry AMD (non-neovascular) is represented by drusen deposition, later evolving into confluent areas of regressed drusen and ultimately in the advanced dry stage presenting as loss of vision associated with retinal pigment epithelium (RPE) atrophy (clinically known as geographic atrophy, GA). Wet AMD (neovascular) is characterized by the leakage of fluid in the sub-RPE and subretinal spaces caused by neovascularization. The overall objective for this study is to design, develop, and evaluate AMD prognosis prediction models that can detect most relevant images containing AMD biomarkers, manage unevenly spaced sequential optical coherence tomography (OCT) images and predict all advanced AMD forms that can help with the interpretation and explainability of computer-aided prognosis models.
Modelling Uncertainty in Medical Reports
Miscommunication of diagnostic uncertainty can deeply affect the quality of treatment a patient receives. A standardized quantification based on the language used in medical reports is a solution for gaining clarity about the amount of uncertainty an author intended to convey. The goal of this project is to first create a dictionary of terms and phrases used in radiology reports that are indications of uncertainty or certainty, and then assign an uncertainty level to a radiology report based on the uncertainty words used in the report. Preliminary results using publically available teaching files repositories of radiology reports such as MIRC and MyPACS show that, by using a dictionary of both certainty and uncertainty descriptors, we can characterize and quantify diagnostic uncertainty of medical reports.
Radiomics-Based Texture Analysis of Idiopathic Pulmonary Fibrosis
Differences in texture features within levels of lungs affected by Idiopathic Pulmonary Fibrosis (IPF) provide a visual representation of genetic mutations and can be used to predictively model patient prognosis. Using CT images, we propose to first quantify texture information in regions of interest (ROIs) in three levels of the lungs along peripheral and internal regions and then use these features to predict genetic mutations at various SNPs of patients using the Receiver Operator Characteristics (ROC) analysis. Our preliminary results using model survival based on texture features and Kaplan-Meier survival curves show that certain thresholds for texture features are suggestive of less favorable IPF prognosis.
Developing a Pipeline for Automatic Evaluation of Grey-White Matter Differentiation in Computed Tomography
This project aims to identify grey-white matter differentiation within the brain on a CT image. This may be a useful metric for predicting a patient’s survival and disability outcomes following a coma after cardiac arrest. To develop a pipeline that does not require patient MRIs, the MNI 152 template and lobe masks in the same standard space can be used to assist us in identifying the regions of interest on CT scans. A comparison of the resulting distributions of pixel values in Hounsfield units between grey and white matter can help find differences between patients with good outcomes and those with poor outcomes.
Applying Topic Modeling Techniques to Detect Trends in Mental Discussion
Mental health disorders may share similar symptoms making differentiating between them for the purposes of diagnosis a difficult task. Bipolar disorder and borderline personality disorder have been shown to fall into this category. Based on the language used by Reddit.com users in forums dedicated to each of these disorders, as well as forums dedicated to the loved ones of individuals with these disorders, the goal is to identify any differences that may increase accuracy of diagnosis. Using natural language processing (NLP) techniques to automatically identify informal topics and patterns in the language extracted from these forums, our preliminary results show that there are significant differences in the language used in these different forums that could be applicable to diagnostic protocols.