If Pulmonary Tuberculosis (PTB) is detected early in a patient, the greater the chances of treating
and curing the disease. Early detection of PTB could result in an overall lower mortality rate.
Detection of PTB is achieved in many ways, for instance, by using tests like the sputum culture
test. The problem is that conducting tests like these can be a lengthy process and takes up precious
time. The best and quickest PTB detection method is viewing the chest X-Ray image (CXR) of
the patient. To make an accurate diagnosis requires a qualified professional Radiologist. Neural
Networks have been around for several years but is only now making ground-breaking
advancements in speech and image processing because of the increased processing power at our
disposal. Artificial intelligence, especially Deep Learning Convolutional Neural Networks
(DLCNN), has the potential to diagnose and detect the disease immediately. If DLCNN can be
used in conjunction with the professional medical institutions, crucial time and effort can be saved.
This project aims to determine and investigate proper methods to identify and detect Pulmonary
Tuberculosis in the patient chest X-Ray images using DLCNN. Detection accuracy and success
form a crucial part of the research. Simulations on an input dataset of infected and healthy patients
are carried out. My research consists of firstly evaluating the colour depth and image resolution of
the input images. The best resolution to use is found to be 64x64. Subsequently, a colour depth of
8 bit is found to be optimal for CXR images. Secondly, building upon the optimal resolution and
colour depth, various image pre-processing techniques are evaluated. In further simulations, the
pre-processed images with the best outcome are used. Thirdly the techniques evaluated are transfer
learning, hyperparameter adjustment and data augmentation. Of these, the best results are obtained
from data augmentation. Fourthly, a proposed hybrid approach. The hybrid method is a mixture
of CAD and DLCNN using only the lung ROI images as training data. Finally, a combination of
the proposed hybrid method, coupled with augmented data and specific hyperparameter
adjustment, is evaluated. Overall, the best result is obtained from the proposed hybrid method
combined with synthetic augmented data and specific hyperparameter adjustment. / Electrical and Mining Engineering
| Identifer | oai:union.ndltd.org:netd.ac.za/oai:union.ndltd.org:unisa/oai:uir.unisa.ac.za:10500/26890 |
| Date | 11 1900 |
| Creators | Norval, Michael John |
| Source Sets | South African National ETD Portal |
| Language | English |
| Detected Language | English |
| Type | Dissertation |
| Format | application/pdf |
Page generated in 0.0023 seconds