Scholarlink

Novel biological and biomedical experiments are performed with synthetic organ-on-chip rather than with animals. These artificially grown tissues (and cells) can effectively simulate human body conditions. It is important to observe the single cell (or tissue) behavior after providing new medication to it. However, the reaction can be visible even after several hours—it will be tiring to look directly at the cell during all this time (even with specialized equipment). Moreover, sometimes the changes are so small that even experienced laboratory worker is not able to see them. Fully automated tools are needed to solve this problem. Their first stage must be segmentation—as we need to point out the appropriate region, where the cell or tissue is visible, in the spatial video. For this aim, we proposed usage of 2D U-Net with Focal loss during training and testing. Worked-out model is trained with 1240 images and evaluated with additional 100 samples. We reached 0.915201 Dice score and 0.848639 mean Intersection over Union (IoU) when comparing to the cells segmented by hand.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Cell Segmentation in Spatial Recordings of Single EHT’s Cardiac Cycle Using Shallow 2D U-Net Trained with Focal Loss

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Maciej Szymkowski

摘要

Novel biological and biomedical experiments are performed with synthetic organ-on-chip rather than with animals. These artificially grown tissues (and cells) can effectively simulate human body conditions. It is important to observe the single cell (or tissue) behavior after providing new medication to it. However, the reaction can be visible even after several hours—it will be tiring to look directly at the cell during all this time (even with specialized equipment). Moreover, sometimes the changes are so small that even experienced laboratory worker is not able to see them. Fully automated tools are needed to solve this problem. Their first stage must be segmentation—as we need to point out the appropriate region, where the cell or tissue is visible, in the spatial video. For this aim, we proposed usage of 2D U-Net with Focal loss during training and testing. Worked-out model is trained with 1240 images and evaluated with additional 100 samples. We reached 0.915201 Dice score and 0.848639 mean Intersection over Union (IoU) when comparing to the cells segmented by hand.