Cellular Architecture on Whole Slide Images Allows the Prediction of Survival in Lung Adenocarcinoma

Pathology is the gold standard for cancer diagnosis. Numerous studies aim to automate the diagnosis based on digital slides, yet its prognostic utilities lack adequate investigation. Besides the inherent difficulties in predicting a patient's prognosis, extracting informative features from gigapixel and heterogeneous whole slide images (WSI) remains an open challenge. We present a computational pipeline that can generate an embedded map to flexibly profile different cell populations' local and global composition and architecture on WSIs. Our approach allows researchers to investigate tumor cells and tumor microenvironment based on these embedded maps of a reasonable size rather than dealing with gigantic WSIs. Here, we applied this pipeline to extract the texture patterns for tumor and immune cell types on the TCGA lung adenocarcinoma dataset. Based on extensive survival modeling, we have demonstrated that by pruning redundant and irrelevant features, the final prediction model has achieved an optimal C-index of 0.70 during testing. Our proof-of-concept study proves that the efficient local-global embedded maps bear valuable information with clinical correlations in lung cancer and potentially in other cancer types, warranting further investigations.