[Defense] Classification of 3D images using ML, Focusing on resource efficiency: memory, “work”, energy
Tuesday, May 18, 2021
11:00 am - 12:00 pm
In
Partial
Fulfillment
of
the
Requirements
for
the
Degree
of
Doctor
of
Philosophy
Nazanin
Beheshti
will
defend
her
proposal
Classification
of
3D
images
using
ML,
Focusing
on
resource
efficiency:
memory,
“work”,
energy
Abstract
In this research we propose the use of Convolutional Neural Networks for the classification of Alzheimer’s disease (AD) patients based on correlation matrices formed by a novel use of functional MRI (fMRI) time series data. Alzheimer’s disease is generally viewed as degraded communication between brain regions termed functional connectivity. In fMRI brain activity is captured from Blood Oxygen Level-Dependent (BOLD) magnetization detected by the MRI scanner. The functional connectivity is inferred from correlations of the observed BOLD signals from different regions of the brain. For AD classification of patients using fMRI the brain is typically observed with a spatial (voxel) resolution of 2 – 4 mm and a sampling rate of 0,5 – 2 Hz for a duration of a five to six minutes creating a 4D data set of a few hundred thousand time series (one per voxel) each of 100 – 200 samples. The brain from a functional perspective is divided into regions captured by what is known as brain atlases. In this research we use the 90 region Automated Anatomic Labeling atlas, AAL-90. The functional connectivity is measured by BOLD signal correlation between regions. Functional connectivity can be assed using stimuli, task-based observations, or without stimuli, resting state observation. The latter is typically used for classification of AD, which is also the basis for AD classification in this research. For the classification the 4D data set is reduced to a 2D data set with one data set per region. The correlation of the data sets between regions is then used for the classification. We form the region data sets in two novel ways: 1) region data sets formed by time-averaged voxel BOLD signals ordered along a Hilbert curve, 2) region data sets formed by concatenation of voxel time series with voxels order along a Hilbert curve. In other studies region data sets are formed by averaging the voxel time series over the voxels in the region. Further, we form regions either around the center points of the AAL-90 regions, or around randomly selected brain coordinates, with region sizes of equal length Hilbert curve segments before and after the AAL-90 center points or randomly selected brain coordinate.
Our initial results show that both our methods for forming region data sets yields better AD classification results than the commonly used spatial averaging of region time series. Further, with inter-region correlation matrices formed from our region data sets, 2-layer CNN yield comparable AD-CN classification to our 4-layer CNN with accuracies in the about 80 – 90% range. For CN-MCI classification, 2-layer CNN yield accuracies in about 90-100% range, and for AD-MCI classification, 2-layer CNN yields accuracies in about 80-90% range. The filter coefficients for the 2-layer CNN requires 253 kiB of memory and the 4-layer CNN 294 kiB. The training time for the 2-Layer CNN and 4-layer CNNs on a test set of 320 subjects required 49 and 94 seconds respectively using Tensorflow.1.14, Python 3.7.3 on a 2.3GHz Intel Core I5-7360U processor with 2 cores and 4 threads, 16 GB 2133 MHz LPDDR3 memory, and macOS Mojave-10.14.6.
Tuesday,
May
18,
2021
11:00AM
-
12:00PM
CT
Online
via
Zoom
Dr. Lennart Johnsson, dissertation advisor
Faculty, students and the general public are invited.
