Histopathology has been used as a very important resource for diagnosis so far. If a lot of By utilizing the accumulated pathological data risk predictions in addition to conventional diagnoses can be made. To predict precise prognostic pathological features, we introduced tumor environment-associated context learning using graph deep learning (TEA-graph), using whole slide images and linked survival information. Using graph neural networks (GNNs), the local pathological features were analyzed with the contextual features reflecting their surrounding environment in whole slide images. TEA-graph improved risk prediction and stratification performance, which would aid both clinicians and patients in real-life settings.
Histopathology of the tumor reflects not only the heterogeneity of cell morphology but also cell-level omics (DNA, RNA, protein, metabolite) information. In particular, image based deep learning has been applied to pathology images to employ this high resolution information and showed it's possibility to perform tasks from cell segmentation to mutation or gene expression, and even origin of metastasis prediction.
However, for these studies to be used in the real world, deep learning models must provide a result in a way that doctors can interpret. For example, if we could interpret the deep learning model that predicts the tumor patient’s survival, we could also know what cell morphology plays an important role in survival of tumor patients. To address this issue, we are developing an explainable deep learning model that not only utilizes the rich information in pathology images, but also provides interpretation of the results. We expect this could be used to find biomarkers related to diagnosis or prognosis.
Recently, we introduced tumor environment associated context learning using graph deep learning (TEA-graph), a graph deep learning model that could predict the prognosis of the tumor patient by considering not only the specific region but also the surrounding environment of the regions. Additionally, we uncovered which cell to cell connections play an important role in survival and suggested a contextual biomarker which showed a bad prognosis in kidney cancer.
