Deep Learning Methods for Classification of Glioma and its Molecular Subtypes

Licentiatavhandling, 2021

Diagnosis and timely treatment play an important role in preventing brain tumor growth. Clinicians are unable to reliably
predict LGG molecular subtypes from magnetic resonance imaging (MRI) without taking biopsy. Accurate diagnosis prior to surgery would be important. Recently, non-invasive classification methods such as deep learning have shown promising outcome in prediction of glioma-subtypes based upon pre-operative brain scans. However, it needs large amount of annotated medical data on tumors. This thesis investigates methods on the problem of data scarcity, specifically for molecular LGG-subtypes.

The focus of this thesis is on two challenges for improving the classification performance of gliomas and its molecular subtypes using MRIs; data augmentation and domain mapping to overcome the lack of data and using data with unavailable GT annotation to tackle the issue of tedious task of manually marking tumor boundaries. Data augmentation includes generating synthetic MR images to enlarge the training data using Generative Adversarial Networks (GANs). Another type of GAN, CycleGAN, is used to enlarge the data size by mapping data from different domains to a target domain. A multi-stream Convolutional Autoencoder (CAE) classifier is proposed with a 2-stage training strategy. To enable MRI data to be used without tumor annotation, ellipse bounding box is proposed that gives comparable classification performance.

The thesis comprises of papers addressing the challenging problems of data scarcity and lacking of tumor annotation. These proposed methods can benefit the future research in bringing machine learning tools into clinical practice for non-invasive diagnostics that would assist surgeons and patients in the shared decision making process.