MPhil Thesis Presentation - Detection and Quantification of Anatomic Hallmarks in Glioma Using Deep Learning on Whole Slide Images

Histopathological analysis is one of the important approaches allowing pathologists to diagnose the grade of glioma and predict prognosis. However, traditional histopathological analysis is time-consuming, subjective, and lacks a clear quantitative standard. In recent years, artificial intelligence-guided histopathology analysis is a new paradigm that enables automatic, accurate, and quantitative analysis of histopathological images. Artificial intelligence methods use data to predict cancer type, grades, and prognosis without requiring a detailed model of the underlying manually engineering cellular features and cancer biological pathways. They accelerate histopathological image analysis by abstractive learning and integrate the information from every single pixel of the image. In this thesis, we begin with an introduction of the basics of artificial intelligence with a focus on applications of histopathological image analysis. We developed a deep-learning guided prediction algorithm to automatically characterize histopathological images and genomics from 2,516 glioma patients. A Deep Convolutional Network (CNN) was developed to classify and quantify necrosis, one of the hallmark patterns of glioblastoma, from patches of histopathology images. The performance of model prediction is comparable viii to that of pathologists. To enhance the interpretability of the model, attention mechanisms in combination with visual explanation showed the model has learned meaningful histologic features for prediction. Model generalization capacity is evaluated on the TCGA dataset, and the result showed Multiview input can boost the performance. To further explore how to take advantage of the model in understanding pathology, a machine learning method was developed based on the feature extracted by the deep learning model to classify histologic patterns. Also, the model can be used to predict patient prognosis with accuracy comparable to the manual histologic-grading model. Moreover, the underlying embedding of the model is used to cluster the necrosis into 2 clusters to distinguish pseudo-palisading necrosis and other subtypes correlated with specific biological pathways. These results demonstrate that automated histopathological image analysis based on deep learning can create a complementary approach to tissue diagnosis that differs from traditional pathology methods. Then we introduced the detailed methods used in this thesis. Last, we discuss other potential pathological questions that can be studied using deep learning in the future.