Exploration of the use of synthetic data generated from DCGANs on COVID-19 detection using CNNs

License: GNU General Public License v2.0

TeX 5.08% Jupyter Notebook 94.57% PureBasic 0.35%

artificial-intelligence cnn-classification cnn-keras covid-19 gans generative-adversarial-network masters masters-thesis letterkenny masters-degree

masters_thesis_automated_detection_of_covid_19_using_gans_and_cnns's Introduction

Masters Thesis

Proposal: Investigating the uses of Generative Adversarial Networks and optimal architecture for use in data generation for training models to diagnose COVID 19

completed as part of Masters of Science in Computing in Artificial Intelligence Research in Atlantic Technological University Letterkenny, under the supervision of Dr. Paul Greaney

What is this project

This project was part of attaining a 60 credit MSc in Artificial Intelligence Research. This project aims to investigate the possibility of using artificially generated data generated by a Generative Adversarial Network(GAN) to train a Convolutional Neural Network(CNN). Due to data shortage there have been many issues regarding the ability of current CNN models to generalize when diagnosing COVID-19 from X-rays when using a CNN, to combat this I have decided to investigate GANs to generate synthetic data to increase a model's abilitiy to generalize. I also investigate transfer learning for automated COVID-19 diagnosis in this project and the results attained when using transfer learning with an augmented dataset.

Structure of Repo

COVID-19 Chest X-Ray - contains images for Covid 19 chest x-ray dataset.
COVID-19_Radiography_Dataset - contains images for COVID-19_Radiography_Dataset
Jupyter Notebooks - contains two subfolders GANs and CNNs which contain CNNs and GANs for each dataset
Literature Review Papers - contains literature reviewed in the thesis
Records of Supervision - Contains formal records of supervision which detail discussions between myself and my supervisor
Thesis Reviews - Reviewed copies of my thesis where changes needed to be made
Thesis - contains unzipped source of thesis
ThesisImages - contains images used in my thesis
xray_dataset_covid19 - contains images for xray_dataset_covid19
ReadMe.md - this file you are reading detailing the project, it's purpose, and explaining it.
Thesis.pdf - Compiled PDF file of Thesis
Thesis.tex.zip - zipped source of Thesis

Sources for datasets:

Radiography dataset: https://www.kaggle.com/datasets/tawsifurrahman/covid19-radiography-database
COVID chest X-ray dataset: https://www.kaggle.com/datasets/bachrr/covid-chest-xray?datasetId=548681
COVID-19 X-Ray dataset: https://www.kaggle.com/datasets/khoongweihao/covid19-xray-dataset-train-test-sets
Extensive COVID-19 dataset: https://data.mendeley.com/datasets/8h65ywd2jr/3

NOTES

The papers within the literature review folder were not written by me but by their respective authors
Many of the images within the thesis images folder were not created by me but belong to their respective creators(cited in thesis)
Datasets were not compiled by me original sources cited in thesis - also above in the sources for datasets section of this ReadMe
Extensive X-Ray dataset was too large to upload to this repo please use the link mentioned in sources for datasets
Due to limitations with Git LFS I was unable to save the GAN models to this repository will include link to drive at a later time.
Link to GAN Models Google Drive: https://drive.google.com/drive/folders/1-uTalRnB5MiZk5jsh4TDGxclDqvoJUaN?usp=sharing
Link to CNN Models Google Drive: https://drive.google.com/drive/folders/1bBrI8bLj5-YWCqaNWRShTPkWvDhM1Ut6?usp=sharing

masters_thesis_automated_detection_of_covid_19_using_gans_and_cnns's People

Contributors

Watchers

Forkers

wiplash742000

masters_thesis_automated_detection_of_covid_19_using_gans_and_cnns's Issues

Start on Basic GAN

Write results section of thesis

Proof-Read Introduction

Optimize All Transfer Learning Models

Need to optimize the transfer learning models for some of the datasets

Investigate Xray COVID Gans

Strange behaviour in X-ray COVID DCGAN it appears that the normal class DCGAN is producing good quality images despite lack of data and Pneumonia model is having issues but may possibly be improved

Start on basic CNN implementation

Rerun CNN models with test split and document

Need to rerun CNN models with test split as this is the only way to ensure objectivity when discussing / analyzing results. Need to document these changes also.

Proof read results section

Documenting the results from training the augmented CNN models need to proof-read to see if results are detailed correctly.

Discuss and compare model's in literature review with augmented models

Start on Literature Review

Work on using GANs in conjunction

Validation accuracy staying static in some CNN models

Validation accuracy staying static in some CNN models need to investigate and find causes

Source Papers for Literature Review

Update paths in GAN notebooks

Need to update paths in GAN notebooks

Rerun Radiography X-Ray DCGAN

Used all data instead of just X-Rays for this particular GAN will rerun and augment again.

Retrain CNN models using all data

Initially used early stopping to increase accuracy but realized this would hurt models ability to generalize to new data.

Start on Implementing different GAN Architectures to train CNN Individually

Document TL models and GAN design for Extensive Dataset

Finish Conclusion Section

Document augmented TL models for X-ray COVID-19 dataset

Need to document results of TL models for X-ray COVID-19 dataset.

Write Implementation section of thesis

X-ray covid 19 CNNs

Before I used a custom split in Keras instead of test / train folders in model. This will have very little effect on initial models but for Augmented models this will be detrimental as I'll essentially be judging the model's accuracy when predicting GAN images. On the other datasets this will have little effect however given the size of this particular set it is likely to be highly detrimental when judging model's accuracy.

Research different CNN architectures to train

Expand on DCGAN results

Detail significance of results in thesis

Make PLT plots less noisy

Will investigate ways to make plt plots less noisy.

Need to rewrite X-ray COVID 19 sections

Noticed the train set was being split using 70 / 30 this is no longer needed as the dataset has a separate directory for train and test

Research transfer learning architectures for CNN

Finish Results Section

Rename some of images in thesis image folder and update CNN model stats

Need to rename some of the image names in thesis image folders as they are a bit confusing and need to update the implementation section of CNN with latest stats.

Fix issue in notebook

Used chatgpt to help write code to filter out the synthetic images, however, when evaluating on the baseline model which previously had an accuracy of >90 % it is now achieving around only 30%. Believe this to be an issue with the way the dataset was filtered.

ultan-kearns / masters_thesis_automated_detection_of_covid_19_using_gans_and_cnns Goto Github PK