Data Augmentation for Breast Cancer Mass Segmentation

In medical imaging, a major limitation of supervised Deep Neural Network is the need of large annotated datasets. Current data augmentation methods, though quite efficient to enhance the performance of deep learning networks, do not include complex transformations. This paper presents a realistic image transformation model mimicking multiple acquisitions obtained from the analysis of a mammography database composed of screening acquisitions with priors. Our transformation model results from the combination of a registration algorithm, an invariant meshing strategy and a reduced model describing motion and local intensity variation in paired images. The extracted data variability was then transferred trough data augmentation to a small database for the training of a deep learning-based segmentation algorithm. Significant improvements are observed compared to usual data augmentation techniques.