Machine Learning Methods for Image Analysis in Medical Applications, from Alzheimer's Disease, Brain Tumors, to Assisted Living

Doctoral thesis, 2020

This thesis investigates artificial intelligence and computer vision techniques with several healthcare-related applications. Among these applications, two most important ones are Alzheimer's disease detection and brain tumor classification.

Alzheimer's disease (AD) is a neural disease that happens mostly in the elderly, and memory loss is its main symptom. We address AD detection since early treatment of AD patients can slow down the progression of AD and reduce the symptoms. An effective method for diagnosing AD has been developed using Magnetic Resonance Images (MRIs). To learn the “features” from normal subjects and AD patients, we have proposed a new “deep learning” method, which automatically learns the changes caused by AD in grey matter (GM), white matter (WM) and cerebrospinal fluid (CSF) regions. These changes in MRIs are used as the features to distinguish AD patients from normal subjects. Our test results from a trained machine system has reached the state-of-the-art performance with more than 90% of AD patients detected. More tests should be conducted in the near future on larger datasets for possible clinical usage.

Another important application is to predict molecular-level types of brain tumor, glioma. Glioma is one of the most common types of original brain tumors. Prediction of glioma subtypes in the molecular level can lead to better treatment strategies. Usually a biopsy is taken from the brain of a patient for determining the tumor subtypes. It is a rather risky step as extracting tissues might lead to the loss of some brain functions such as hearing and vision. Hence, the process is usually performed while a patient is kept fully awake. By exploring deep learning methods to learn the molecular-level glioma subtype information from patients using their MRIs and biopsy information, our proposed method is shown promising to predict the glioma subtype by purely using MRIs from a new patient, without requiring a biopsy. The performance is rather encouraging, and it has reached the state-of-the-art (accuracy over 85% for one subtype, IDH mutation/wild-type prediction) though far from ideal. Further improvement of prediction performance is needed by improving the methods and by using a much larger number of MRIs to train the machine. We hope some improved deep learning methods can, in the near future, provide medical doctors with better semi-automatic ways for diagnosing brain tumors non-invasively.