Deep Learning-Based Identification of Microglial Cellular Populations in Glioblastoma Patients.

Glioblastomas (GBMs) are primary brain malignancies with poor overall survival; however, disease sequelae following primary treatment is highly variable in nature-leading to variation in therapy response and overall survival. Recent discoveries characterizing the underlying GBM microenvironment evidence dynamic responses in both residential and recruited immune cell populations which may in fact drive these clinical observations. To validate these finding and evaluate the potential role of immune cells in differential survival, we analyzed recurrent GBM RNAseq data from the Chinese Glioma Genome Atlas and subsequently performed unsupervised clustering through k-means to define 3 sub-clusters of recurrent GBM with differential overall survival (p = 0.019). Subsequently, to evaluate the role of immune cells in this paradigm, we applied Weighted Gene Correlation Network analysis to our findings which indicate modulation of immune response as a key driver of this differential clinical pattern. Translating these findings to actionable clinical utilization in neuropathology is however challenged by poor signal-to-noise ratios during immuno-staining and complex arborization of microglial processes. We thus present a convolutional neural network workflow using a UNET framework in python which can mitigate these challenges and successfully annotate diagnostic pathology slides from real glioblastoma patients treated at Ohio State with 80% accuracy-with attribution of lowered accuracy due to annotation error. Using this approach, we plan to subsequently annotate additional histology slides and utilize the morphologic variation within our segmented microglial population to discern post-surgical response patterns in glioblastoma patients.