Date of Award

January 2017

Document Type

Thesis

Degree Name

Medical Doctor (MD)

Department

Medicine

First Advisor

Jeff Geschwind

Second Advisor

James Duncan

Abstract

The aim of this thesis is to present results from two original research projects that involve computational approaches to clinical radiology with a special emphasis on hepatocellular carcinoma (HCC). The projects were undertaken sequentially and may be viewed as components of a common pipeline. They focus on the automated calculation of response-to-treatment criteria for patients with HCC who undergo transarterial chemoembolization (TACE). TACE is an endovascular procedure involving the local delivery of a chemotherapeutic or embolic agent into vessels supplying a tumor.

Response-to-treatment criteria are measures that radiologists use to quantify how well patients respond to therapy. Traditional response criteria are time-consuming to calculate and suffer from inter-operator variability. They are not uniformly performed despite an abundance of evidence demonstrating their value. The chief objection to their use is the time required for their manual calculation.

A fully automated method for response-to-treatment calculations would remove the time burden on radiologists and facilitate standardized reporting on HCC lesions. Successfull automation requires automatic liver and tumor segmentations and automatic metric calculation. Given the former (i.e. segmentations), the latter becomes straightforward.

The first project demonstrates a simple heuristic to fully automate the quantitative European Association for the Study of the Liver (qEASL) response criterion, assuming liver and tumor segmentations are provided. Our method automates parenchymal region-of- interest (ROI) selection and shows result equivalent to manual ROI selection achieved by the current qEASL standard.

The second project applies recent developments in deep learning for computer vision to the problem of automatic liver and tumor segmentation. We train a deep convolutional neural network classifier to detect liver and tumor margins in contrast-enhanced MR imaging.

These projects represent an end-to-end solution to tumor response criteria. They provide radiologists a straightforward method for response calculations and produce consistent, reproducible results.

Comments

This thesis is restricted to Yale network users only. This thesis is permanently embargoed from public release.

Share

COinS