Deep learning for glioma histopathology image classification

Artificial Intelligence (AI) based methods have been successfully implemented in many tumour studies to assist pathologists and medical professionals in fast and accurate decision making. Glioma is one of the most common type of brain cancer in adults. At present, a patient‘s cancer treatment is decided by the evaluation of hematoxylin and eosin (H&E) stained histopathology slides by a trained and experienced pathologist. This is a time-consuming task which depends on the availability of the experts and the results are often subjected to observer bias, which could impact subsequent treatment decisions. Digital pathology helps on automating the process, saving time and providing a faster decision-making process. However, the large size and different input formats of the whole slide images in association with the heterogeneity of the glioma tissues make the analysis a challenging task. This research on glioma histopathology images focuses on developing AI based techniques to successfully classify gliomas. The methods use glioma whole slide images (WSIs) from The Cancer Genome Atlas (TCGA) dataset to implement automated classification of the WSIs. The methods classify glioma images according to the 5th edition of the World Health Organization (WHO) Classification of Central Nervous System Tumors (2021), into the histologic categories: astrocytoma, IDHmutant, oligodendroglioma, IDHmutant and 1p/19qcodeleted and glioblastoma, IDHwildtype. Two different deep neural networks (VGG19 and ResNet50) and two classification methods (ensemble and multiclass) are implemented. Glioma images stain normalised using traditional and generative adversarial network (GAN) based methods are used to study the impact of stain normalisation on glioma classification task. Data augmentation and Kfold cross validation techniques are also used in this research. Explainable AI techniques used visualises the features used by the deep learning models in classifying the glioma WSIs. The experimental results of the classification tasks demonstrate high performance (> 80%), which shows that the proposed method can support examination and differential diagnosis of glioma histopathology images. The contributions of this research include, (1) an automated AI based classification pipeline for glioma histopathology images, (2) role and impact of stain normalisation in histopathology image analysis, (3) explainability of the predictions made by the deep learning models. The outcomes of this research contribute an important role in building novel, automated methods for glioma histopathology image classification, which can augment the decisions of pathologists in tumour detection and diagnosis.