BR-GAN: Bilateral Residual Generating Adversarial Network for Mammogram Classification

Abstract

Mammogram malignancy classification with only image-level annotations is challenging due to a lack of lesion annotations. If we can generate the healthy version of the diseased data, we can easily explore the lesion features. An intuitive idea of such generation is to use existing Cycle-GAN based methods. They achieve the healthy generation regarding healthy images as reference domain, while maintaining the original content by cycle consistency mechanism. However, healthy mammogram patterns are diverse which may lead to uncertain generations. Moreover, the back translation from healthy to the original remains an ill-posed problem due to lack of lesion information. To address these problems, we propose a novel model called bilateral residual generating adversarial network(BR-GAN). We use the Cycle-GAN as a basic framework while regarding the contralateral as generation reference based on the bilateral symmetry prior. To address the ill-posed back translation problem, we propose a residual-preserved mechanism to try to preserve the lesion features from the original features. The generated features and the original features are aggregated for further classification. BR-GAN outperforms current state-of-the-art methods on INBreast and in-house datasets.