MRI Data Augmentation Using GANs (MTHE 493 Thesis)

Challenge

Machine learning in healthcare has a lot of potential uses, one of them being faster and more accurate diagnosis. Medical data collection with the sole use of building datasets is too expensive, which often leaves researchers with small datasets. In addition to being small, they often have issues with class imbalance and do not reflect real-world data.

Approach

The team and I applied Generative Adversarial Networks (GANs) to address the issues of data scarcity and class imbalance by exploring the use of synthetic data’s ability to augment and replace small datasets. We were able to produce GANs that closely approximated the distribution of the dataset, and determine the optimal parameters and best loss function.

Contribution

Designed and implemented the GAN architecture in Python with Tensorflow and Keras
Implemented a class of parameterized loss functions
Trained and tested models, generated synthetic images for reports
Wrote sections of the final thesis, including the background, methodology, and future work

Figure: A real MRI scan (left) next to a generated MRI scan (right), both with tumors.