Quantitative chemical imaging: A top-down systems pathology approach to predict colon cancer patient survival

Colon cancer is the second deadliest cancer, affecting the quality of life in older patients. Prognosis is useful in developing an informed disease management strategy, which can improve mortality as well as patient comfort. Morphometric assessment provides diagnosis, grade, and stage information. However, it is subjective, requires multi-step sample processing, and annotations by pathologists. In addition, morphometric techniques offer minimal molecular information that can be crucial in determining prognosis. The interaction of the tumor with its surrounding stroma, comprised of several biomolecular factors and cells is a critical determinant of the behavior of the disease. To evaluate this interaction objectively, we need biomolecular profiling in spatially specific context. In this work, we achieved this by analyzing tissue microarrays using infrared spectroscopic imaging. We developed supervised classification algorithms that were used to reliably segment colon tissue into histological components, including differentiation of normal and desmoplastic stroma. Thus, infrared spectroscopic imaging enabled us to map the stromal changes around the tumor. This supervised classification achieved >0.90 area under the curve of the receiver operating characteristic curve for pixel level classification. Using these maps, we sought to define evaluation criteria to assess the segmented colon images to determine prognosis. We measured the interaction of tumor with the surrounding stroma containing activated fibroblast in the form of mathematical functions that took into account the structure of tumor and the prevalence of reactive stroma. Using these functions, we found that the interaction effect of large tumor size in the presence of a high density of activated fibroblasts provided patients with worse outcome. The overall 6-year probability of survival in patient groups that were classified as “low-risk” was 0.73 whereas in patients that were “high-risk” was 0.54 at p-value <0.0003. Remarkably, the risk score defined in this work was independent of patient risk assessed by stage and grade of the tumor. Thus, objective evaluation of prognosis, which adds to the current clinical regimen, was achieved by a completely automated analysis of unstained patient tissue to determine the risk of 6-year death. In this work, we demonstrate that quantitative chemical imaging using infrared spectroscopic imaging is an effective method to measure tumor-tumor microenvironment interactions. As a top-down systems pathology approach, our work integrated morphometry based spatial constraints and biochemistry based stromal changes to identify markers that gave us mechanistic insights into the tumor behavior. Our work shows that while the tumor microenvironment changes are prognostic, an interaction model that takes into account both the extent of microenvironment modifications, as well as the tumor morphology, is a better predictor of prognosis. Finally, we also developed automated tumor grade determination using deep learning based infrared image analysis. Thus, the computational models developed in this work provide an objective, processing-free and automated way to predict tumor behavior.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Bioengineering

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2019

Author dc:creator

• Tiwari, Saumya

Contributors dc:contributor

• Bhargava, Rohit
• Kajdacsy-Balla, André
• Pan, Dipanjan
• Sweedler, Jonathan

• Infrared Spectroscopic imaging, cancer, outcome, survival, machine learning

Statement dc:rights

• Copyright 2018 Saumya Tiwari

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/102938

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/102938