Computer Vision News - February 2022

49 METGAN: Generative Tumour Inpainting tells us. “ With this method, we can image the same specimen in multiple channels to highlight different aspects of the organism. We have an autofluorescence channel at the base to better image the anatomy. We impose a label that we use as a ground truth label to obtain a contrast channel – what we call in our paper the cancer channel – where we highlight objects of interest. In our case, it was metastases . These are small tumors that are spreading throughout the body. ” This paper uses a generative adversarial network approach to solve the problem of needing high-quality labeled ground truth data to develop robust segmentation methods . The team found state-of-the-art methods did not prove useful or sufficient for generating the labeled data required. They sought to develop a new method to generate more diverse data and improve the segmentation for novel specimens. “ The imaging method we use in our lab is called light sheet microscopy , ” Izabela Inference. In order to generate realistic annotated data, we can use a combination of a real background and a pre-determined, user-specified label. Within our pipeline, we train two generators and two discriminators in a cycle-consistent manner. The generator that is creating the synthetic tumour domain images is further constrained by a pre-trained segmentor.