Date of Award

1-1-2019

Document Type

Thesis

Degree Name

Medical Doctor (MD)

Department

Medicine

First Advisor

James Duncan

Abstract

Advances in technology have provided humanity with tools to combat one of its greatest scourges, cancer. In the late 19th century, the discovery of X-radiation paved way for the development of tools that can enable visualization of structures inside the body and to assess tissues non-invasively. This allowed for earlier detection of a wide range of pathologies including neoplastic disease, as well as provide a platform for the development of image-guided cancer therapies. More recently, advances in computer vision and “artificial intelligence” have allowed for detailed analysis of radiographic characteristics of diseased tissue and the identification of clinically significant features not apparent to the human observer, particularly in oncologic imaging.

The goal of this thesis is to contribute to the technical development of novel cancer imaging and radiomics approaches using artificial neural networks, and to highlight imaging based diagnostic and therapeutic technologies that are already impacting cancer care. The current trajectory of imaging and technological adoption in the diagnosis and management of neoplastic disease is readily apparent in this work.

The first aim is achieved with two exploratory studies assessing the utility of artificial intelligence enabled radiomics in decoding tumor phenotypes and for prognostic analysis in lung cancer. We used deep-learning based approaches to predict tumor histology in early stage non‑small cell lung cancer (NSCLC) using non-invasive computed tomography (CT) data. We also find that convolutional neural network (CNN)- derived CT-radiomics features have prognostic value and can be used to stratify early stage NSCLC patients into long-term and short‑term survival groups. Deep-learning approaches have emerged as robust alternatives to traditional statistical pattern recognition algorithms including Bayesian methods and probabilistic graphical models for image analysis, thereby presenting viable alternatives for image interpretation and classification.

The second aim was achieved by reviewing a multi-centered study assessing the impact of MRI-guided laser thermal therapy on outcomes in patients with brain metastases failing stereotactic radiosurgery. We also present a case in which a patient with an atypical presentation of hemolytic anemia found to have a neuroendocrine tumor benefited from somatostatin receptor-based imaging techniques in determining the etiology and extent of his disease.

Improvements due to novel computational approaches for image processing and analysis may help accelerate the technical capacity for image-guided treatments, as well as improve image reconstruction in some of the widely utilized modalities for cancer imaging. Furthermore, deep-learning and data science may help us gain new insights from outcomes in large patient datasets. Similarly, current patient experiences and results from clinical studies may help inform future directions of technical development. Therefore, this thesis serves as an example of how two or more domains of expertise, clinical and technical, can work synergistically to either answer new questions, or old questions using new techniques. All work presented here is based on scientific publications in which the candidate was first author, or on technical and computational approaches developed by the candidate and presented at national and international meetings.

Comments

This thesis is restricted to Yale network users only. It will be made publicly available on 07/15/2021.

Share

COinS