細胞次結構影像辨識系統

A Subcellular Structure Images Recognition System

李玫憶(Mei-Yi Li)

指導教授 ： 蔡育秀

中原大學/工學院/醫學工程研究所/碩士(2005年)

https://doi.org/10.6840/cycu200500778

摘要

人類基因體計畫完成後，生物學家發現擁有正確的DNA排列序，並不足以確保基因會表現出原本欲表現之功能。預測基因功能方法很多，從蛋白質分佈於細胞內之位置預測功能就是其中之ㄧ。目前，對細胞標記螢光，用螢光顯微鏡觀察並拍攝細胞內部之次結構影像，已經完全自動化。然而，辨識蛋白質分佈之次結構影像仰賴研究人員長期視覺觀察經驗的累積，使得自動化大量建立細胞內蛋白質分佈位置預測功能受到限制。

本研究主要目的為建立一套可以跨越細胞種類限制與不同顯微影像擷取方法之細胞次結構影像辨識系統，在於縮短研究人員觀察實驗影像時間、提高研究效率，並幫助經驗不足之人員辨識次結構種類。八種典型細胞次結構標定基因與各色活體螢光蛋白表現載體相結合，轉染至CHO、Vero細胞中表現，用兩種不同倒立式螢光顯微鏡取得此八種螢光細胞次結構顯微影像。細胞次結構影像首先進行影像前處理（如：形態濾波)，突顯次結構外觀特性後，對影像擷取幾何與材質特徵。再利用逐步鑑別統計分析，篩選出對辨識次結構影像具鑑別力之特徵集合，利用篩選後之數值特徵資料建立決策樹。

最後，用CHO和Vero兩種類細胞之次結構影像做系統測試。從系統結果發現，本研究對845張CHO細胞影像樣本平均辨識率為86.2%，而317張Vero細胞影像樣本，其平均辨識率也達82.7%。綜合兩類細胞影像，整體系統辨識率為85.3%，已有輔助實驗人員辨識參考之價值。

關鍵字

影像辨識 ； 蛋白質分佈 ； 決策樹 ； 細胞次結構影像

並列摘要

After the completion of Human Genome Project, scientists find that right gene sequence can not ensure the genes will perform original expressed function. In fact, the gene production, protein, affect the function of the cell. Understanding the mechanism and function for protein is more important. There are lots of methods to predict the function of protein. One of them is recognition of the protein localizations to predict their functions. In the lab, scientists using immunofluorescent staining and fluorescence microscopy to acquire images of protein localization in the cells. However, recognition of cell images relies on human observation, and thus became the obstacle in classifying subcellular protein localizations by junior researchers. To construct an automated image recognition system will shorten post-analysis process. Several research groups using specific features extraction to recognize subcellular protein localizations with high accuracy. These recognition systems are restricted to images of single particular cell line acquired by the specific imaging system.

The study is to establish a recognition system of subcellular structures of various cell types acquired by different imaging systems. Standard marker genes of eight subcellular structures (actin, ER, golgi, peroxisome, mitochondria, microtubule, nucleus,nucleolus ) fused with fluorescence protein are expressed in CHO and Vero cell, then images taken by fluorescence microscopy as training and test materials for system. First, we use image preprocess technology, for examples: morphological filter, to emphasize the shapes of subcellular structures. Extracted numeric geometric and Harlick texture features from processed images were applied for, stepwise discriminant analysis to select statistically significant features of their ability to separate the eight classes of subcellular structures. Finally, use selected numeric features to build the decision tree.

845 images of CHO cells images and 319 images of Vero cells were used to test the performance of the system. Results showed that the system can recognize subcellular structures of CHO cell images with 86.2% accuracy, 82.7% for Vero cell and 85.3% accuracy for a mixture of CHO cell images and Vero cell images. Our system is capable of recognizing subcellular structures of various cell types acquired by different images.

並列關鍵字

Subcellular structure ； Protein localization ； Image recognition ； Decision tree

參考文獻

[12]熊家誠，"自動化螢光顯微影像之次細胞結構辨識"，中原大學醫學工程研究所碩士論文，民國九十三年。

[23]黃金聰、史天元，"以碎形為基礎的內插模式與雙線性、立方迴旋內插的比較"，航測及遙測學刊3（4）：39-52。

[5]Boland,M.V., Markey,M.K. and Murphy,R.F. "Automated recognition of patterns characteristic of subcellular structures in fluorescence microscopy images." Cytometry, 33, 366–375. SMC 8, 630–632. 1998.

[7]Boland MV, Murphy RF. "A neural network classifier capable of recognizing the patterns of all major subcellular structures in fluorescence microscope images of HeLa cells." Bioinformatics, 2001;17:in press.

[8]K. Huang, M. Velliste, and R.F. Murphy, "Feature Reduction for Improved Recognition of Subcellular Location Patterns in Fluorescence Microscope Images," Proc SPIE, 4962: 307-318, 2003.