Deep Learning Glioma Grading with the Tumor Microenvironment Analysis Protocol for A Comprehensive Learning, Discovering, and Quantifying Microenvironmental Features

Abstract

AbstractGliomas are primary brain tumors that arise from neural stem cells or glial precursors. Diagnosis of glioma is based on histological evaluation of pathological cell features and molecular markers. Gliomas are infiltrated by myeloid cells that accumulate preferentially in malignant tumors and their abundance inversely correlates with survival, which is of interest for cancer immunotherapies. To avoid time-consuming and laborious manual examination of the images, a deep learning approach for automatic multiclass classification of tumor grades was proposed. Importantly, as an alternative way of investigating characteristics of brain tumor grades, we implemented a protocol for learning, discovering, and quantifying tumor microenvironment elements on our glioma dataset. Using only single-stained biopsies we derived characteristic differentiating tumor microenvironment phenotypic neighborhoods. A challenge of the study was given by a small sample size of human leukocyte antigen stained on glioma tissue microarrays dataset - 203 images from 5 classes - and imbalanced data distribution. This has been addressed by image augmentation of the underrepresented classes. For this glioma multiclass classification task, a residual neural network architecture has been adapted. On the validation set the average accuracy was 0.72 when the model was trained from scratch, and 0.85 with the pre-trained model. Moreover, the tumor microenvironment analysis suggested a relevant role of the myeloid cells and their accumulation to characterize glioma grades. This promising approach can be used as an additional diagnostic tool to improve assessment during intra-operative examination or sub-typing tissues for treatment selection, despite the challenges caused by the difficult dataset. We present here the distributions and visualizations of extracted tumor inter-dependencies.Graphical abstractHighlightsResearch highlight 1: We demonstrate that the ResNet-18 architecture with simple data augmentation trained in 10-fold cross-validation performs the multiclass classification relatively well even with a small imbalanced dataset with a high degree of similarities between classes.Research highlight 2: After supervised subtyping of the tumor, we investigated the usefulness of the protocol for discovery and learning tumor microenvironment elements for the same task. The protocol designed for deriving new biomarkers based on multiplex stained histological samples proved the ability to detect features characteristic of malignant tumors based only on single target stained tissue microarrays. We propose further studies on this topic can help in formulating specific criteria for improvements in diagnosis of gliomas, allowing to avoid the necessity of conducting advanced histopathological analysis or complementing genetic testing of tumor samples.