Chest XRAY GAN REsearch at UNE

     During my senior year of high school, I conducted independent research at the University of New England (UNE) with the help of Dr. Sylvain Jaume, focused on generating high-quality synthetic chest X-ray images using machine learning. My project utilized a Conditional Generative Adversarial Network (CGAN) and a Super-Resolution Generative Adversarial Network (SRGAN) to improve the realism and resolution of generated medical images.
     The CGAN was trained to produce chest X-rays conditioned on specific pathologies, ensuring that the model could be given a label to create high accuracy images for different diseases (No Findings, Mass, Pneumonia). Meanwhile, the SRGAN was employed to enhance the
resolution of these images, making them
suitable for diagnostic and research
applications. My research aimed to
address data scarcity in medical imaging
by providing a reliable method for
generating synthetic X-rays that could
aid in deep learning model training and
radiology education. One notable part of
this research was its one-of-a-kind
structure to allow for low memory use.
Historically, GANs use an incredible
amount of memory, however this model
was designed in a way allowing it to be
ran on a Nvidia GTX 1050Ti
     I had the opportunity to present my
findings at the Summer Undergraduate
Research Experience (SURE) Symposium
at UNE alongside Dr. Jaume, where I
discussed the development process,
challenges encountered, and the
potential applications of my work.
The image on the right captures a
moment from my presentation,
showcasing my engagement in research
at an advanced level while still in high
school. Additionally, a more detailed
research paper, available on the right
(currently not published), provides an in-depth breakdown of my methodology, experimental results, and future directions. To demonstrate the effectiveness of my model, I have also included 50 generated chest X-ray images (and a csv file with their diagnosis), which highlight the quality and variability achieved through my CGAN/SRGAN framework.
     This project required extensive knowledge of deep learning, particularly in training and optimizing Generative Adversarial Networks. I worked with TensorFlow and Keras to build and fine-tune the CGAN and SRGAN models, adjusting hyperparameters to balance generator and discriminator performance. Additionally, I implemented data preprocessing techniques to standardize medical images, including histogram equalization and contrast adjustments, ensuring consistency in the training set. These technical skills allowed me to produce a model capable of generating realistic chest X-rays while maintaining crucial pathological features.
   This research exemplifies my ability to integrate machine learning, image processing, and data science to solve real-world problems. It also highlights my experience in academic research, from conducting literature reviews to presenting findings at a professional symposium. Through this project, I demonstrated technical expertise in GAN architecture, proficiency in Python and deep learning frameworks, and an ability to communicate complex concepts effectively. The combination of research, engineering, and presentation skills showcased in this work reflects my readiness to tackle AI challenges in both academic and applied settings.