Deep learning system for brain tumor grading classification

The project aims to classify about 300 high- and low-grade glioma cases from MICCAI - BRATS 2015 Challenge Dataset. Each record has T1 MRI, T1 contrast-enhanced MRI, T2 MRI, and T2 Flair MRI. We use a convolutional neural network architecture for grade classification of each case. The main hallma...

Full description

Saved in:
Bibliographic Details
Main Author: Nhan, Tran Ho Chi
Other Authors: Lin Zhiping
Format: Final Year Project
Language:English
Published: 2017
Subjects:
Online Access:http://hdl.handle.net/10356/72015
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:The project aims to classify about 300 high- and low-grade glioma cases from MICCAI - BRATS 2015 Challenge Dataset. Each record has T1 MRI, T1 contrast-enhanced MRI, T2 MRI, and T2 Flair MRI. We use a convolutional neural network architecture for grade classification of each case. The main hallmark of the architecture is the utilization of spatial information of a tumor by using 3D convolutional layers with 3D kernels and 3D max pool layers. Since GPU memory is an essential factor of quickly training a neural network, the number of filters in each layer is carefully reduced and tested so that it won’t cause the network to lose its ability to characterize important features of the volume dataset. Furthermore, the size of each input volume is also reduced in a way that still maintains the crucial tumor features. This was achieved by cropping only the part of brain which contains the tumor, and forwarding it through 5 paths of 3D convolutional layers and 3D max pool layers. To objectively maintain the trustworthiness of the network, the small amount of data set - 223 high-grade cases and 57 low-grade cases was overcome by cross-validation, i.e. we split the data set to multiple train-test sets so that the network’s performance can be seen on all test cases. Finally, stochastic gradient descent was implemented at the input phase of the network by loading small batches of random data from the training set instead of expensively shuffling the whole training set for many iterations. The quality of the network is periodically assessed using the test set for each train-test run. The result shows an improvement with the 3D CNN and deeper architecture compared with the previous 2D CNN network.