Registries provide a valuable tool for cancer research and for enabling decision support systems. However, populating cancer registries with information from medical records can be a tedious and bottlenecking process. This thesis presents an annotator system that automatically extracts data elements from lung cancer radiology reports to populate a lung cancer registry. Annotators systems such as this utilize natural language processing (NLP) techniques to locate concepts from a text source. A web-based framework that wraps the annotator system into a graphical user interface for researchers and clinicians is also discussed.

Lung cancer is the leading cause of cancer-related death in both sexes. Large, randomized clinical trials have demonstrated that low-dose computed tomography screening reduces mortality from lung cancer, as opposed to chest X-ray or no screening, when participants adhere to follow-up recommendations. However, low adherence rates in post-trial clinical lung cancer screening (LCS) programs have been reported across the United States (US). Low adherence to LCS in real-world clinical practice diminishes the mortality benefit of annual screening derived from clinical trials. Thus far, limited studies have examined factors affecting the patient’s decision (not) to adhere to screening guidelines. This dissertation examines the factors that may predict patient non-adherence to LCS recommendations. First, we performed a systematic review and meta-analysis of 24 studies published between 2014 and 2020 that mentioned adherence to Lung-RADS recommendations, identifying factors contributing to adherence rates. Second, using the Carter-Harris conceptual model, which enumerates psychosocial variables (e.g., smoking stigma, cancer fear, cancer fatalism) related to LCS participation and adherence, we examined these variables’ availability and completeness in our medical records. Next, using a subset of variables where data were consistently available, we identified factors of non-adherence over multiple screening time points using logistic regression and mixed effects models. Lastly, we used statistical and machine learning-based methods to examine how well we could predict patient non-adherence using longitudinal data across three screening time points. This dissertation advances our understanding of factors contributing to patient non-adherence to LCS recommendations. It provides a basis for identifying patient groups that could benefit from individualized interventions to improve LCS adherence.

The reprogrammed energy metabolism and the dysfunctional vascular network of tumors create a hypoxic and acidic microenvironment, which is related to various malignant properties of cancer and poor patient prognosis. We have developed an amine chemical exchange saturation transfer (CEST) sequence with spin-and-gradient echo (SAGE) echo-planar imaging (EPI) readout to evaluate tumor acidity and hypoxia in human gliomas simultaneously. Amine CEST provided pH-sensitivity through labeling the endogenous amine protons that undergo chemical exchange with water protons, with a pH-dependent exchange rate. On the other hand, the reversible transverse relaxation rate quantified using the multi-echo EPI readout reflects oxygen extraction through sensitivity to paramagnetic deoxyhemoglobin.

This dissertation focused on developing and validating this novel dual-function imaging technique, mainly from three aspects: the technical development and validation, the biological validation, and the clinical validation of the proposed pH- and oxygen-sensitive CEST-SAGE-EPI technique in human gliomas. We have developed a new post-processing method for improved $B_0$ correction. A customized CEST physical phantom was designed and developed with validated temporal stability. We also evaluated the CEST contrast variability in healthy volunteers and the normal-appearing contralateral brain regions in glioma patients. The proposed pH- and oxygen-sensitive imaging biomarkers showed significant correlations with the tumor cell metabolomics features and MRI-guided biopsy tissue biomarkers, which validated the biological bases of the imaging biomarkers. Additionally, we have examined the association between tumor acidity with tumor vascularity, as measured by perfusion MRI. Lastly, we investigated the clinical usefulness of the biomarkers to characterize different glioma genotypes, predict patient prognosis, and monitor treatment responses.

In summary, this dissertation demonstrated that the novel dual-function pH- and oxygen-sensitive imaging technique reflects the abnormal metabolism in glioma patients and has the potential to provide clinical values for patient diagnosis, prognosis, and treatment efficacy assessment.

Stroke is the fifth leading cause of death in the United States, with approximately 795,000 new cases each year. The goal of stroke treatment is to rescue salvageable tissue by reperfusion therapy. Clinical trials have shown that intravenous tissue plasminogen activator (IV tPA) and clot retrieval devices are effective treatments for recanalizing occluded blood vessels. However, determining an optimal stroke treatment plan is not a straightforward decision because it involves different factors, such as patient risk of hemorrhage and penumbra size. The relationships between these factors and patient outcomes are not clearly understood. This dissertation attempts to overcome these challenges by developing machine learning and deep learning models for acute ischemic stroke clinical and imaging outcome classification. A novel deep learning model was first proposed using source perfusion imaging to predict voxel-wise tissue outcome. The model architecture is designed to include contralateral patches to improve the feature learning process. Second, an end-to-end machine learning approach was developed to classify stroke onset time, which is a major clinical variable in selecting patients for IV tPA treatments. The approach combines baseline descriptive features and deep features to improve stroke onset time classification using machine learning models. Third, a bi-input convolutional neural network was developed for perfusion parameter estimation. This model lays a foundation to estimate perfusion parameters using pattern recognition techniques. Finally, a machine learning model trained with a balanced data set was developed for acute stroke patient outcome prediction. Rigorous experiments and results have shown the effectiveness of these proposed methods. This dissertation describes methods that lead to better understanding of stroke imaging, which lays the foundation to offer decision-making guidance for clinicians providing acute stroke intervention treatments.

Currently, low-dose computed tomography (LDCT) is the only recommended screening test for patients who are at high risk of lung cancer. However, the cost-benefit analysis of LDCT must be weighed against the high number of false positives, radiation exposure, unnecessary procedures, and the associated patient distress as a result of the aforementioned possibilities. Sequential decision making models such as the partially observable Markov decision process (POMDP) have seen success in making recommendations in clinical applications such as lung cancer screening. Enabled by the availability of longitudinal datasets that track patient health over time, these models make predictions toward long-term health outcomes. A key challenge in lung cancer screening is the balancing between true positives and false positives, that is, maximizing true positives while minimizing false positives. This dissertation attempts to address this challenge by leveraging a variety of techniques toward optimizing decision making over time. First, the modularized POMDP (modPOMDP) framework is developed to account for temporal variations within a POMDP model. Each time point is optimized separately to ensure "earlier" detection by maximizing positive predictions over the entire screening duration. Second, a two-part model framework (modPOMDP2) is developed to differentiate true and false positive predictions from each other. This method combines classic machine learning (ML) techniques and modPOMDP to maintain true positives while decreasing false positives. Third, the validity of these approaches is demonstrated in an external testing dataset.

As the volume of biomedical literature increases, it can be challenging for clinicians to stay up-to-date on this massive store of knowledge. Graphical summarization systems condense knowledge into a more tractable form via "concept maps" -- networks of nodes (concepts) and edges (relations between concepts). In existing graphical summarization systems, the context of the extracted relations (such as study design and study population) is omitted. However, context is crucial for capturing the full meaning of a relation. With context, the user may pose more detailed queries than those accommodated by traditional, context-free maps.

This dissertation describes Casama, a system for creating "contextualized semantic maps" to represent the current state of scientific knowledge in the domain of non-small cell lung cancer (NSCLC). A formalism for contextualized semantic maps is presented, including targeted relations, study design context, and study population context. An annotated gold standard conforming to this representation is produced, and methods for extracting these contexts are developed. Contextualized semantic maps are evaluated in an information retrieval task and a summarization usability study, showing significant improvement over PubMed and SemRep.