使用 U-Net 以分割三維細胞影像內小鼠胚胎幹細胞

Segmentation for mouse embryonic stem cells on 3D images by U-Net

陳俊瑜(CHUN YU CHAN)

指導教授 ： 蔡明達

中原大學/電機資訊學院/資訊工程學系/碩士(2022年)

https://doi.org/10.6840/cycu202201521

摘要

三維細胞容積中心以單細胞分辨率進行細胞分割，可為基礎生物學和應用研究使用。然而，3D 共聚焦影像存在信號干擾比低、螢光響應弱、沿影像疊加方向分辨率不足等問題。使用影像處理方法或與幾何處理方法，很難從細胞容積中的緊密或接觸細胞中分離出單細胞。最近，3D 深度學習方法已經被醫學使用來避免影像和幾何處理中繁瑣的參數設置，但仍然不容易分割出靠近或接觸的單個細胞。因此 本文提出了一種 2D U-Net 以高精度和快速計算以分割細胞區域。另外，透過計算幾何的方式分割細胞區域與 2D U-Net的分割結果比較，通過結合 3D U-Net 來檢測容積中單個細胞的中心，可以實現更好的 3D 細胞影像和緊密或接觸細胞容積中的單細胞分割。並將會以三種不同架構進行三維分割方式，來實現三維共軛焦細胞螢光影像的視覺化。

關鍵字

三維機器學習 ； 三維細胞影像 ； 單細胞辨識 ； U-Net ； 共聚 焦顯微鏡影像

並列摘要

The 3D cell center detection performs cell segmentation at single-cell resolution and can be used for basic biological and applied research. However, 3D confocal images have problems such as low signal-to-interference ratio, weak fluorescence response, and insufficient resolution along the image stacking direction. With image processing methods or geometric processing methods, it is difficult to separate single cells from close or touching cells in a volume. Recently, 3D deep learning methods have been used in medicine to avoid cumbersome parameter settings in image and geometry processing, but it is still not easy to segment individual cells that are close or touching. This paper proposes a 2D U-Net to segment cellular regions with high accuracy and fast computation. In addition, the segmentation of cell regions by computational geometry is compared with the segmentation results of 2D U-Net. By combining 3D U-Net to detect the center of single cells in the volume, it is possible to achieve better single-cell segmentation for 3D cell images with close or contact cells. And three different architectures are used for volume segmentation to find the best segmentation method to improve the visualization of 3D confocal fluorescence images.

並列關鍵字

3D machine learning ； 3D cell segmentation ； single cell recognition ； U-Net ； confocal microscopy cell image

參考文獻

[1] C.L. Baker and M.F. Pera, “Capturing totipotent stem cells,” Cell Stem Cell, 22(1), pp. 25-34, 2018

[2] Rodriguez-Terrones D., Gaume X., Ishiuchi T., Weiss A., Kopp A., Kruse K., Penning A., Vaquerizas J.M., Brino L. and Torres-Padilla M.E., “A molecular roadmap for the emergence of early-embryonic-like cells in culture,” Nature Genetics, vol. 50, pp.106-119, 2018.

[3] X. Lou, M. Kang, P. Xenopoulos, S. Mun˜oz-Descalzo and A.K. Hadjantonakis, “A rapid and efficient 2D/3D nuclear segmentation method for analysis of early mouse embryo and stem cell image data,” Stem Cell Reports, vol. 2, pp. 382-397, 2014.

[4] Chang Y.H., Yokota Y., Abe K.,Tsai M.D. and Chu S.L., 2021. Automatic three-dimensional segmentation of mouse embryonic stem cell nuclei with multiple channels of confocal fluorescence images. Journal of Microscopy, 281(1),57-75.

[5] Falk T., et al. 2019. U-Net – Deep Learning for Cell Counting, Detection, and Morphometry. Nature Methods, 16, 67–70.