The goal of this project is to synthesize natural looking sick mammography images that will expand the database of a mammography classifier, and by that improve its success rates. Deep learning classifying networks require a large amount of data to get good results.
In the medical field there is a lack of tagged, anonymous sick-patient’s data.
Therefore, we synthesized sick images using deep learning techniques, in this work - CNN, in order to enlarge the classifier’s training data.
In the first stage we took the simplest approach for creating synthetic sick images. We implanted real tumors from sick images into healthy mammography images, and trained a simple CNN to fill the edges. We encountered a problem with this method when the tumor and the healthy image had different brightness. In these cases, the results were non-realistic.
In the second stage, we created a fully synthesized tumor on a healthy image. We used a u-net architecture, multiple loss functions and transfer learning techniques to fill “tumor shaped” holes with tumor-like tissue in healthy images, and by that to create sick images that are completely synthetic.
Finally, when adding 1000 of the synthetic images to a classifier, its AUC criterion improved by 5%.