The application of deep learning to nucleus images for early cancer diagnostics

Abstract

Cancer remains a major concern for patients and early diagnosis can go a long way in treating patients. Current cancer diagnosis usually involves a pathologist looking at tissue slices of patients for specific features associated with cancer prognosis such as nuclear morphometric measures. However, early diagnosis remains a major challenge. Recent studies have shown that changes in fibroblast nuclei play a critical role in the early development of cancer. In addition, it is crucial that computational models are capable of justifying themselves when used in critical decisions such as diagnosing a patient with cancer. In this thesis, we use machine learning techniques on two dimensional nuclei images to show that computational models are capable of presenting human interpretable features as a means of justifying themselves. In addition, we use machine learning techniques on volumetric images of nuclei of cells in a co-culture model that represents the cancer tissue microenvironment to study changes the fibroblasts undergo. These studies pave the way for various approaches to early disease diagnosis.