根據粒線體及細胞核顯微影像利用深度學習分類細胞週期

Deep-Learning Classification of Cell Cycle Phases Based on Mitochondrial and Nucleus Microscopic Images

黃麒戎(Chi-Jung Huang)

指導教授 ： 魏安祺

國立臺灣大學/電機資訊學院/生醫電子與資訊學研究所/碩士(2022年)

https://doi.org/10.6342/NTU202200281

摘要

細胞在生長過程中可分為四個階段，屬於間期的G1、S、G2以及分裂期 (M)，不同階段的細胞除了外型變化外，內部的胞器如細胞核、粒腺體型態也會隨生長有所不同。要觀察細胞週期常用的方法為流式細胞術，其可以分析螢光強度判斷細胞位在哪個週期。現今的流式細胞儀也可進行影像拍攝，但其缺點為儀器拍攝時放大倍率及影像解析度不足，無法輕易分辨細胞較細小的胞器。在本篇論文中，我們利用fluorescence ubiquitination-based cell cycle indicator (FUCCI)及MitoTrackerTM對AC16心肌細胞的細胞週期及粒線體標定，並使用能拍出高倍率及高解析度影像的雷射共軛焦顯微鏡來拍攝時間序列上細胞變化的影像。得到的細胞穿透光及螢光影像則用於分類細胞位於哪一週期階段。在拍攝細胞影像時，為避免使用過多雷射通道造成的光毒性問題，我們不直接標定細胞核，而是利用U-net對細胞核螢光影像進行預測。最後我們將預測出的細胞核影像與拍攝的顯微鏡影像一同放入卷積神經網路ResNet及MobilenetV2訓練預測細胞位於哪個週期階段，在精確度上相比只放入螢光及穿透光影像有所提升。整體而言使用預測的螢光影像即可幫助提高分類週期的準確性，這個方法可以降低事前準備螢光染色的時間，更能讓觀察顏色有限的螢光顯微技術多了彈性調整的空間，讓研究者能觀察其他胞器在不同週期中的變化。

關鍵字

細胞週期 ； FUCCI ； 螢光影像預測 ； U-net ； ResNet ； MobilenetV2

並列摘要

The cell cycle is the main process that regulates cell growth and development. During the different cell cycle phases, in addition to changes in the cell’s shape, the morphology of organelles, such as the nucleus and mitochondria, changes. Flow cytometry is the most commonly used method to classify cell cycle phases by measuring the intensity of fluorescence from nucleus DNA. While imaging flow cytometers can be used to acquire cell images in different cell cycle phases, the magnification and resolution are not sufficient to record microscopic changes in the organelles. In this thesis, we labeled the cell cycle phases and mitochondria of the human cardiomyocyte AC16 cell line using Fluorescent Ubiquitination-based Cell Cycle Indicator (FUCCI) and MitoTrackerTM, respectively. High-resolution, time-lapsed images were acquired using a confocal microscope with a high-magnification objective to monitor cell growth and division. To avoid phototoxicity, we trained a U-net-based model that can predict fluorescent-labeled images of the cell nucleus from transmitted light images. The predicted cell nucleus images, along with the transmitted light cell images and fluorescent-labeled mitochondria images, were taken into the convolutional neural networks ResNet and MobilenetV2 to train in order to predict which cycle stage the cell was in. FUCCI, which labels the G1 phase and S/G2/M phases, was used as ground truth in the model training.
Compared with only fluorescent-labeled or transmitted light images, convolutional neural networks provide increased prediction accuracy, with additional predicted nucleus images from transmitted light images, resulting in an improved classification of cell cycle phases. The deep learning method proposed in this thesis, which uses transmitted light images to predict cell cycle phases, can reduce the time required for sample preparation with fluorescent labeling, is more flexible than fluorescence microscopy, and allows researchers to observe the changes in organelles in different cycle phases.

並列關鍵字

Cell Cycle ； FUCCI ； Prediction of Fluorescence Image ； U-net ； ResNet ； MobilenetV2

參考文獻

[1] L. H. Hartwell and T. A. Weinert, “Checkpoints: controls that ensure the order of cell cycle events.,” Science, vol. 246, no. 4930, pp. 629–634, Nov. 1989.

[2] A. Murray, “Cell cycle checkpoints.,” Curr. Opin. Cell Biol., vol. 6, no. 6, pp. 872–876, Dec. 1994.

[3] E. A. Nigg, “Cyclin-dependent protein kinases: key regulators of the eukaryotic cell cycle.,” Bioessays, vol. 17, no. 6, pp. 471–480, Jun. 1995.

[4] F. Girard, U. Strausfeld, A. Fernandez, and N. J. Lamb, “Cyclin A is required for the onset of DNA replication in mammalian fibroblasts.,” Cell, vol. 67, no. 6, pp. 1169–1179, Dec. 1991.

[5] M. Ohtsubo, A. M. Theodoras, J. Schumacher, J. M. Roberts, and M. Pagano, “Human cyclin E, a nuclear protein essential for the G1-to-S phase transition.,” Mol. Cell. Biol., vol. 15, no. 5, pp. 2612–2624, May 1995.