- 學位論文
不同深度學習演算法應用於細胞影像分割之比較與大腸桿菌質體分離實例分析
Comparison of Different Deep Learning Algorithms Applied to Cell Image Segmentation and Case Analysis of Plasmid Partition in Escherichia coli
摘要
活體細胞縮時攝影可以產生大量的數據，隨之而來的問題則是如何將影像中的細胞分割出來。在傳統影像處裡方面，Otsu演算法與分水嶺演算法 (Watershed Algorithm)是兩種常見的將影像二值化的方法。對於細胞影像分割，處在細胞較稠密或是緊鄰的狀態下，傳統影像處裡無法達到完美的細胞分割結果。因此，我們將深度學習中的電腦視覺應用於細胞影像分割，選擇了SuperSegger、Unet、Mask R-CNN這三種模型進行細胞分割的比較與分析。在這三種模型中，表現最出色的為Unet，並以此作為實例分析中影像處裡的基礎。 在大腸桿菌質體分佈實例分析的部分，過去的研究顯示多套數質體並沒有類似低套數質體主動分離的機制，且有可能由轉錄或是轉譯所導致質體群聚的現象，因此多套數質體的穩定維持機制並不明確。我們以螢光抑制操作系統標記多套數的CoE1衍生質體，透過長時間的細胞縮時攝影實驗，搭配影像分割模型Unet，統計在有無抑制轉錄的情況下，子代細胞分配到親代細胞質體的比率。發現在抑制轉錄的情況下，質體分配的比率更集中在成功機率為1/2的二項分布寬度內。我們記錄了不同的養菌條件下，質體數量在細胞中成長的變化情形。盡管細胞在洋菜膠上仍然會分裂，但在細胞中質體數量比較多的情況時，細胞是傾向不消耗能量去複製質體。在細胞極性的統計實驗當中，我們發現新端細胞比舊端細胞有更容易獲得質體的趨勢，這個現象可能來自於細胞中類核在空間上的分布不均勻，且在第二次細胞分裂前，質體的聚集出現在靠近新端附近的頻率可能是比較高的，因此新端細胞會更有機會分配到較多的質體。
並列摘要
Live cell time-lapse microscope can generate a lot of data, and the problem that comes with it is how to segment the cells in the image. In traditional image processing, Otsu algorithm and Watershed Algorithm are two common methods to binarize images. For cell image segmentation, when the cells are dense or in close, the traditional image processing methods cannot achieve perfect cell segmentation results. Therefore, we applied deep learning to cell image segmentation, and selected three models include SuperSegger, Unet, and Mask R-CNN to compare with each other and analyze their performance on cell segmentation. Among these three models, the best model is Unet, which is used as the basis for image processing in the following case analysis. In the part of the case analysis of E. coli plasmid partition, there is evidence that high-copy number plasmid does not have a mechanism similar to the active partition system of low-copy number plasmid. In addition, for high-copy number plasmid, there is an existence of aggregation which may be caused by transcription or translation. The mechanism of plasmid maintenance is still unclear. We labeled high-copy number of CoE1-derivative plasmid with the fluorescent repressor-operator systems. Through Live cell time-lapse microscope experiments, with the cell segmentation model Unet, we recorded the distribution of plasmid inheritance of daughter cells with or without transcription inhibition. It was found that in the case of transcription inhibition, the ratio of plasmid inheritance was more concentrated within the width of the binomial distribution with a success probability of 1/2. We recorded the growth rate of plasmid under different bacterial growth conditions. Although the cells division still occur on agar, when the number of plasmids in the cells is relatively large, the cells tend not to replicate the plasmid. In the section of cell polarity, we found that the new pole cells tend to obtain plasmids more easily than the old pole cells. This phenomenon may be caused by the uneven spatial distribution of nucleoids in the cells and the frequency of plasmid cluster in the new pole may be relatively high. Thus, the new pole cell has a better chance to inherit more plasmids.
參考文獻
延伸閱讀
- 陳鈴宛（2010）。微陣列多時間點基因表達的貝氏分析及應用在桿狀病毒株於不同宿主的比較〔碩士論文，國立清華大學〕。華藝線上圖書館。https://doi.org/10.6843/NTHU.2010.00069
- Chen, Y. C., Chen, S. J., Chang, M. C., & Chen, T. L. (2005). 以不同方法自大腸桿菌細胞間質釋放重組蛋白之比較. Journal of the Chinese Institute of Chemical Engineers, 36(5), 527-532. https://doi.org/10.6967/JCICE.200509.0527
- Ko, C. L., & Wang, F. S. (2007). 重組大腸桿菌在批式與饋料批式培養之菌體和胞內蛋白質的線上估測. Journal of the Chinese Institute of Chemical Engineers, 38(3&4), 197-203. https://www.airitilibrary.com/Article/Detail?DocID=03681653-200707-38-3&4-197-203-a
- 郭金龍、周如文、郭振鵬、孫永年（2020）。Automatic Identification of Acid-Fast Bacilli With Sputum Smear Microscopy Using Artificial Intelligence Deep Learning Method。北市醫學雜誌，()，1-13。https://doi.org/10.6200/TCMJ.202201/PP.0012
- 郭金龍、周如文、郭振鵬、孫永年（2022）。Automatic Identification of Acid-Fast Bacilli with Sputum Smear Microscopy Using Artificial Intelligence Deep Learning Method。北市醫學雜誌，19(3)，278-290。https://doi.org/10.6200/TCMJ.202209_19(3).0006