Integrated multiparametric deep spatial phenotyping of mouse models of lung adenocarcinoma

Abstract

In this thesis, I developed computational pipelines and algorithms that use high dimensional biomarker imaging data to predict features of tumor tissues taken from a genetically engineered mouse model (GEMM) of lung adenocarcinoma. I extracted biomarker expression levels and morphological, textural, and spatial motifs of single cells from the imaging data and used these features to train algorithms to predict tumor histologic grade, a measure correlated with the malignant potential of a tumor. The algorithm predictions were evaluated through comparison to a validated deep learning model. The random forest algorithm achieved a 72% accuracy classifying cells as belonging to a non-tumor, grade 1, grade 2, or grade 3 region and achieved a 87% accuracy classifying cells as belonging to a tumor or non-tumor region. A combination of biomarker, morphological, textural, and spatial features generated models that performed better than any single group of markers by itself; spatial features in particular significantly improved model performance.