Computer Vision News - June 2019

Every month, Computer Vision News reviews a research paper from our field. This month we have chosen two. Here is "Multi- Resolution Networks for Semantic Segmentation in Whole Slide Images” . We are indebted to the authors ( Feng Gu, Nikolay Burlutskiy, Mats Andersson, and Lena Kajland Wilén at ContextVision ), for allowing us to use their images. Their paper is found here . Analysis of histological images is one of the most commonly used tools in the diagnosis and research of a large variety of pathological conditions. Over the last few years, the field of digital pathology has undergone a significant progress thanks to the development of new scanning technologies alongside with the development in the field of computer vision. These newfound advances opened new powerful computational opportunities yielding higher performance levels in terms of accuracy and speed. Whole slide imaging (WSI) is a technology allowing the acquisition of high- resolution digital images representing entire tissue slices scanned from glass slides. WSI contains multi-resolution information organized in a pyramid structure, which allows spatial navigation along multiple magnification levels. This multi-resolution information is used by pathologists and researchers to characterize tissues in levels ranging from sub-cell to multi cell complexes. In some cases, the ability to examine the tissue in multiple resolutions is crucial. For example, in cancer diagnosis, both local information such as regularity of cell shapes and cell density as well as contextual information such as global tissue structure, are highly important for achieving an accurate evaluation. The digitation of histopathological images enables the development of automatic tools assisting in their analysis. Deep neural networks, and specifically Convolutional Neural Networks (CNN) became a gold standard in multiple image processing tasks. Due to their large size, WSIs are commonly cropped into small patches and in some cases also down sampled during the learning and prediction processes. These procedures lead to the loss of either contextual or local information and may damage the learning capability and prediction accuracy in turn. To address this issue, Feng Gu et al. have designed a U-net based multi-resolution network (MRN) allowing the use of multiple resolutions during the learning and prediction processes. The classic U-Net architecture, widely used for the segmentation of histological images, consists of an “encoder” in which the feature maps are down-sampled, a “decoder” up-sampling them back and skip-connections concatenating 10 Research by Dorin Yael Computer Vision News Research “ This approach yields superior results comparing to those gained using a single resolution by the classic U-Net ”