École Nationale Supérieure de
Techniques Avancées

Abstract

Deep learning models allowed the achievement of many breakthroughs in computer vision. Armed with convolutional network, models can classify images, locate objects and generate synthetic images. Training efficient networks requires a large amount of data which is a challenge for medical images. In addition malignant cases represent a fractions of collected data. In this work, we study the application of Generative Adversarial network on mammography and tomography patches. We generate high fidelity labeled images using Progressive Growing GANs at 512x512 resolution. However, they were not beneficial when used for training. PGGAN training set contained 26,000 patches. We also training a cross modality image-to-image translation models that transforms annotated mammography patches to tomography patches. The synthetic patches are realistic and capture the malignancy of the source image. By fine tuning a CNN classification network on generated tomography patches we managed improved its performance.