Computer Vision News - July 2019

4 Computer Vision News Nuclei Detection and Segmentation Every month, Computer Vision News reviews a successful project. Our main purpose is to show how diverse image processing techniques contribute to solving technical challenges and real world constraints. This month we review challenges and solutions in a Medical Imaging project by RSIP Vision : Challenges in Nuclei Detection and Segmentation . Project Many diagnostic and research processes require an accurate evaluation of the types, shapes and numbers of cells in a specific tissue area. Thus, our goal in this project was to detect and segment single cell nuclei in slices of tissue . To achieve this goal, we had to overcome multiple challenges: the very high variability between the properties of nuclei belonging to different cell types including shapes, sizes, color intensities and textures which vary within and between tissues. In addition, sometimes cells are clustered and need an expert eye to separate them into single cells with definite borders. Another difficulty is due to occlusion, when one cell is partially hidden by another cell. These make it very challenging to build a single process which detects them all in every single slide. There is no single set of parameters which covers them all. Therefore, a smart solution was needed . We separated the project into two steps: the first step focused on detecting the cell, while the second step focused on the segmentation of cells detected during the first step. To do this, we used deep learning techniques which specifically adapted U-Net with different ground truths. RSIP Vision also performed the annotation for this project: this was a difficult task due to the large amounts and high density of nuclei in every single slice. Observations are very subjective as well: specifically in clusters of cells, inter- and intra- observer variability is particularly high. Of course, the accuracy of the model is dependent on the accuracy, number and richness of the annotations. For this project we annotated a few thousands of nuclei using different approaches. The first step of processing in our solution was cell detection . Once the nuclei are detected, the cell segmentation task started, with the goal of segmenting each cell separately (instance segmentation). Once this is done automatically, accurately and fast, the output of this task can be used both for diagnostic and for research purposes. The standardization that we were able to apply with our solution solves the problem of variability in the observations. The fact that this is done automatically reduces the human time needed and enables to answer more questions. If the physician needs to make a diagnostic based on cell detection, classification or count, they will be able to do it faster and better , since it will be independent from human individual observations. by Dorin Yael “… a smart solution was needed!”