76 Chapter 3 to get an accurate description of what is happening in the culture. Counting and measuring the size of multiple organoids per dome is a manual process that can take several hours depending on the density of the cultures. These measurements are highly influenced by factors such as tiredness, personal interpretation, etc. This can lead to low reproducibility between measurements. Brightfield images of the cultures provide a good opportunity to get a broader impression of the health/state of the culture. These organoids can then be manually counted and measured using different softwares. To do so, scientists need to manually select the area of the organoid of interest and trace it in order to get an approximation to the size. Organoid cultures can vary in density, showing as little as a few organoids to as many as a hundred. Another difficulty of tracking organoid’s growth in time, is the need to image the cultures repeatedly (i.e. every 24 hours) and finding the same organoid at every time point in order to track its growth rate. Overall, this task is highly time-consuming and has a very high person-dependent variability. OrgaQuant13 is an open-source implementation that addresses this issue. However, this tool does not perform a pixel-wise segmentation of organoids. Instead, it uses a bounding box to detect an organoid. Thus, the area of an organoids cannot be computed accurately. Additionally, OrgaQuant does not allow for user interventions, and does not relate organoids from different stack images of a Z-stack image. Another group14 has also published annotated organoid images, in which each organoid is surrounded by a bounding box created by experts. However, these data can only be accessed upon contacting the authors. Due to the use of bounding boxes, the correct area of each organoid cannot be computed. This is important since the growth characteristics are crucial for organoid research. Even though they trained a deep neural network to detect and track organoids, this model is not publicly available. In order to overcome these problems, we have extended the OrganelX e-Science service15 to segment and analyze organoids culture. This system allows researchers to upload their brightfield images in order to quantify the number of organoids present in the image as well as to measure their area. The system allows researchers to calculate the average area of multiple organoids of a culture or only of those of interest. It also offers the possibility to select multiple organoids of interest in different Z-stacks of the dome (Figure 1A). Moreover, the system indicates the organoids that are found in different stacks. Another application would be to track organoids in time using images acquired at different time points. The contributions of this work are presented as follows: (1) Design of a deeplearning algorithm and processing pipeline dedicated to organoid-image
RkJQdWJsaXNoZXIy MTk4NDMw