Not stated in source 2017 Needs review

Deep learning methods for MRI brain image analysis: 3D convolutional neural networks for Alzheimer's disease detection and brain tumor classification

Visit original source

Abstract

In this thesis, we investigate deep learning methods for medical image analysis, in particular brain disease detection and classification.We attempt to solve two problems, namely (1) Alzheimer's disease (AD) detection and (2) Brain tumor (BT) classification into high-grade and low-grade gliomas.Efficient and accurate diagnosis of AD is essential for initiating treatment at an early stage, increasing the chances of postponing the irreversible neurodegenerative effects of the disease.To this end, many studies have investigated image processing and machine learning methods for computer-aided-diagnosis (CAD) of AD.In the first part of our work, we continue along this line, and propose a simple 3D CNN architecture for feature extraction and classification of pre-processed T1-weighted MRI brain images for AD detection.The dataset used for training and evaluation consists of 340 subjects, and was obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI).The resulting test performance on unseen data is a classification accuracy of 98.74%, with a sensitivity (AD detection rate) of 100% which is comparable with the current state of the art.In this thesis, we thoroughly investigate the effect of some parameters involved in the CNN training process, such as dropout probability and learning rate.We also show the impact of data pre-processing on the end performance, by comparing several different pre-processing pipelines.For the second part of our work, namely brain tumor classification, we employ a similar method using CNNs, but using a different architecture.The dataset was obtained from the MICCAI BRaTS competition 2017 and consists of 285 subjects.Our preliminary study on brain tumor classification has shown promising results with a classification accuracy of 85.96%, but further work is required for a more in-depth study of the topic.

Keywords

Not available

Specialisations

Similar theses