Fully generated mammogram patch dataset using CycleGAN with de-identification texture analysis

To protect the privacy of patients whose medical information is shared for learning purposes, the US Department of Health and Human Services requires the removal of Personal Health Indicators (PHIs) that may be used to link medical data back to the original patient. The removal of these PHIs, or further obfuscation from the original patient, is known as de-identification. While many of these PHIs can only be found in text data, organizations often place further restrictions on their datasets. Useful medical images are often not publicly shared due to the complexity of ensuring that images cannot be linked to the original medical patient. The de-identification of data, particularly medical data, can be a complicated task that is only becoming more burdensome as larger, more complex datasets become more prevalent. Generative adversarial networks (GANs) are commonly used for image augmentation to produce higher quality results from image classifiers. If GANs could be used to generate sample medical images that cannot be traced back to any singular human patient, this synthetic dataset could be shared easily without restriction or chance of harm to connected patients. Our research aims to take a real medical dataset and transform it via GAN technology to create an entirely synthetic dataset that can be considered de-identified. We then train a deep learning classifier and test that classifier using real data, using this accuracy as a metric for the efficacy of our synthetic dataset. Finally, a gray level co-occurrence matrix-based texture analysis is used to analyze image similarity between the generated images and their real counterparts, offering a quantitative measurement of the de-identification. This project uses the existing, open-source PyTorch CycleGAN implementation. It also uses the MNIST digit dataset for proof of concept of the overall method. Patches extracted from the DDSM dataset are used as a test medical dataset.