運用數;學形態學方法於 cDNA 微陣列;之影像處理
cDNA Microarray Images Processing Using Mathematical
Morphology Mehtods
中文摘要 cDNA 微陣列技術已成為基因體學技術的當紅功能，面對 cDNA 微陣列以快速且 同時檢測上萬個基因的表現模式所產生的大量數據，一個全自動化或接近全自動 化的微陣列影像處理分析軟體是必要的；但是由於 cDNA 微陣列本身的特性， 想要以一種方法來解決各種 cDNA 微陣列影像分析所面臨的問題並不容易。目 前市面上微陣列影像處理分析軟體無論是否為商業或免費軟體大多還需要人為 介入部分操作以及設定各種參數，此外這些軟體所使用的影像處理方法並不一定 完全適用於各種實驗影像狀況。 本研究之目的，在於建立一個接近自動化處理 cDNA 微陣列影像的方法流程； 本論文以數學形態學分水嶺演算法作為影像分割的基礎，配合 Radon Transform 影像旋轉校正、影像投影尋找網格位置、以及各種數學形態學運算子，利用這些 技術來實現自動化處理微陣列影像的目標。 本研究中分別嘗試兩個方法流程，第一個方法按照傳統處理步驟，以尋找網格為 主，依序進行網格定位、分割、擷取強度動作，所有雜交點都限制在個別網格範 圍內；第二個方法則以尋找分割區域為主，將影像中所有有螢光反應的區域分割 出來後再以網格位置判斷分割區域是否為所尋求的雜交點部份。簡言之，兩個方 法最大差別在於，尋求的雜交點區域大小時是否有被限制在網格範圍內。 為了評估本研究試驗的兩種處理方法，採用網路上公開使用的已被分析過的 cDNA 微陣列影像以兩種方法分別進行處理，並同時利用常用軟體(GenePix、 ScanAlyze)等現有的 cDNA 微陣列分析軟體處理，或原先影像所附的分析過的資 料，將研究方法得到的前景強度數據與現有軟體資料數據進行相關性分析並繪製 散佈圖，計算研究方法的數據與使用軟體的數據的線性迴歸模式。結果分析顯 示，無論是否指定網格，本研究使用的兩種方法處理結果與利用 ScanAlyze 以及 GenePix 處理得到的數據有著很高的相關性，且分析結果極為相近；而不指定網 格的處理方法又較指定網格的方法對這些常用軟體的結果擁有更高的相關性與 更好的迴歸模式。在背景強度方面，使用成對樣本 T 檢定進行檢測，結果顯示 本研究用來計算背景強度的方法較 GenePix 的背景強度方法能得到更低的背景 強度評估。 實驗結果證明本研究所提出兩種處理 cDNA 微陣列影像的方法，與常用軟體如 GenePix、ScanAlyze 能得到極為相近的結果；此外，本研究提出的方法確實可以 達到全自動分析 cDNA 微陣列影像的目標。 英文摘要cDNA microarray is widely used to identify and quantify different gene-expression for large-scale analysis. In order to extract microarray data precisely from microarray images, robust image processing is indispensable. Most microarray image processing methods are still semi-automatic because of the variations between different arrayers and spots properties. Besides, in a research always use many cDNA microarrays, and a cDNA microarray allow the monitoring of expressions for tens of thousands of genes simultaneously.An automatic method to resolve these cDNA microarray images quickly and accurately is very important.
This study attempts to propose an automatic spot detection method using
mathematical morphology technology. Watershed transform is the main technique for our spot automatic detection method. Furthermore, Radon transform is used to correct the oblique images and projection technique is used to find the spot grids. To avoid the over-segment problem brought on the sensitive to noises of gradient image used in watershed algorithm, internal and external markers for allocating watershed
immersion positions were also used to reduce the number of image segments.
Two methods were tested in this research. The first method was designed as follows: first, find the spot grids; then, segment spot from a grid; last, calculate foreground and background intensity. The second method was designed to segment fluorescent region from image directly. Next, calculate foreground and background intensity of every segment regiong. Last, compare every segment region with spot grids to screen out non-spot regions.
We compare our experimental results with those obtained from the popular software ScanAlyze and GenePix. Correlation coefficient and linear-regression statistical methods where employed to illustrate the relationship between our methods’
foreground intensity and popular software foreground intensity. Paired-samples T test was used to test the difference between our methods’ background intensity and popular software background intensity. The result showed that the foreground intensities of our methods and popular software have high correlation and the regression model and the coefficient of determination between our methods and popular software reached good suitability, especially in the second method. The background intensities of our methods were really lower than popular software background intensities.
We present fully automatic methods for microarray image processing. It showed that our methods achieved automatic spot detection on non-supervised environment. The advantage of our methods is that it does not have any restrictions for the shape of spots. The methods automatically locate subarray grids, individual spots, and calculate distance between spots or blocks requiring no user identification of any image parameters.
