الفهرس 
Thesis ObjectiveOnly 14 pages are availabe for public view
 |  | from | 138 |  |  |
| | | from | 138 | | |
Abstract
Chest X-ray is one of the main medical imaging modalities for diagnosing lung
diseases. Computer-aided diagnosis (CAD) can be utilized to automatically diagnose
some lung diseases using artificial intelligence techniques. To assist radiologists during
the diagnosis, this thesis aims to apply various deep learning approaches to classify
various thorax diseases automatically. First, pre-trained models are used to accomplish
the diagnostic task of 14 different thoracic diseases in the ChestX-ray14 dataset.
ResNet50 achieved the best performance of multi-label classification of normal and 14
different lung diseases with AUC of 0.911 and F1-score of 0.66.
Second, an automated Machine Learning (AutoML) model is presented to find the
effective architecture and set the suitable hyperparameter of the model for x-ray image
classification. The results of experiments showed that the performance of the AutoML
model achieved an accuracy of 97.8%, F1-score of 97.23%, and AUC of 97% for
pneumonia vs. normal, while the performance viral vs. bacterial pneumonia achieved an
accuracy of 91.4%, F1-score of 90.5%, and AUC of 89.7%.
Then, a hybrid model of convolution neural network (CNN) and long short-term
memory (LSTM) network is proposed to diagnose and classify pneumonia diseases in
pediatric x-ray images. The evaluation of the results showed that the proposed deep
classifiers achieved an accuracy of 98.6% and an AUC of 99.9% for normal versus
pneumonia classification, while the obtained accuracy and AUC scores for bacterial
versus viral classification were 92.3% and 94.5%, respectively.
In comparison to some presented models in the previous studies, AutoML was used to
automatically generate a deep learning model to improve pneumonia detection and
achieve remarkable testing accuracy. In addition, the hybrid model of CNN and LSTM
achieved relatively better performance to assist radiologists in diagnosing x-ray images
than the other adopted ones.
Finally, to improve the performance of the diagnoses process x-ray based, a CNN
model is proposed with an additional segmentation pre-processing step. The x-ray
images are cropped using horizontal and vertical histograms to crop the images and
remove the unrelated regions of the lungs. Then, the CNN model and the hybrid one
are used to diagnose the x-ray images. The proposed CNN model performed an
accuracy of 99.3% and AUC of 100%.