Train small, generate big: Synthesis of colorectal cancer histology images

Abstract

The construction of large tissue images is a challenging task in the field of generative modeling of histopathology images. Such synthetic images can be used for development and evaluation of various types of deep learning methods. However, memory and computational processing requirements limit the sizes of image constructed using neural generative models. To tackle this, we propose a conditional generative adversarial network framework that learns to generate and stitch small patches to construct large tissue image tiles while preserving global morphological characteristics. The key novelty of the proposed scheme is that it can be used to generate tiles larger than those used for training with high fidelity. Our evaluation of the Colorectal Adenocarcinoma Gland (CRAG) dataset shows that the proposed model can generate large tissue tiles that exhibit realistic morphological tissue features including glands appearance, nuclear structure, and stromal architecture. Our experimental results also show that the proposed model can be effectively used for evaluation of image segmentation models as well.