Stability and Robustness of Radiomic Features Due to Volumetric Uncertainty in Pancreatic Cancer

Loading...
Thumbnail Image

Authors

Wong, Jeffrey Y.

Issue Date

2020-08-14

Type

Dissertation

Language

en_US

Keywords

Research Projects

Organizational Units

Journal Issue

Alternative Title

Abstract

Soaring interest in radiomics research seeks to identify quantitative imaging features for enhanced clinical decision support. The radiomics analysis workflow involves image acquisition followed by volume segmentation. Features such as tumor shape or texture (such as how 'smooth' or 'irregular' the tissue appears) can be extracted within the volume and used to discover patterns and correlations that could serve as prognostic markers or build predictive models. Segmentation is a critical and preliminary step to data integration and model building. Variations in segmentation between observers lead to interobserver variability and is a known phenomenon. Some studies have investigated such effects for a few cancer sites such as lung, brain, breast, and prostate. However, tumor delineation uncertainty highly depends on the tumor site, and pancreatic adenocarcinoma (PDAC) is among the most difficult to delineate and therefore more likely to be affected. No studies to date have investigated the impact of interobserver variability on the stability and robustness of CT derived radiomic features for PDAC. Furthermore, while radiomics research is predominantly conducted by radiologists and radiation oncologists, no studies have investigated whether there exists an interdisciplinary difference in terms of segmentation in the context of radiomics. This study investigated this issue of interobserver and interdisciplinary variation in segmentation and quantified its effect on the reproducibility of radiomic features. A high degree of variation was observed between radiology and radiation oncology contours. This was also reflected in the robustness of radiomic features. Simpler features such as first order and shape features demonstrated moderate to good reproducibilty, while complex textural and higher order features showed low reproducibility. These results stress the need for radiomics research to carefully consider the quality of the data used in data integration and model building to prevent unrepresentative results.

Description

Citation

Publisher

Creighton University

License

Copyright is retained by the Author. A non-exclusive distribution right is granted to Creighton University and to ProQuest following the publishing model selected above.

Journal

Volume

Issue

PubMed ID

DOI

ISSN

EISSN