Reti Generative Avversarie per Sintesi e Aumento di Immagini Istopatologiche

Generative Adversarial Neural Networks (GANs) have recently achieved state-of-art performance for image synthesis and have been used to generate high-quality images of faces, animals, and 3D scenes. In this work, we explore the use of GANs in the medical computer vision field, where datasets are usually more scarce and require expert annotations. We demonstrate that GANs can be used to generate high-detailed images of medical objects, such as high-resolution microscopical organ tissue scans. In particular, we focus on the UniToPatho dataset, which contains Hematoxylin and Eosin stained (H&E) colorectal histopathological scans, and evaluate the performance of synthetic images used as an augmentation approach for classification tasks. To further improve the classification accuracy, we propose a new image-to-image StyleGAN-based model that is able to transform segmentation masks of nuclei into realistic tissues. Our model drastically improves the generation quality over previous approaches for the UniToPatho dataset and, most importantly, allows to exploit prior medical knowledge about tissue morphology, by manually guiding the synthesis process for a more precise augmentation. ​