圓形霍氏轉換於cDNA微陣列晶片影像分析之應用

Abstract

本研究利用微陣列影像中雜交點為圓形的特性，以圓形霍氏轉換為中心法則對雜交點進行搜尋。由於圓形霍氏轉換對於所分析的影像具有極高的敏感度，故前處理中雜訊的去除佔了很重要的角色。在本研究當中，首先採用直方圖等化的技術，將影像對比拉大，使前景與背景較易被分離出來，再運用Otsu的二元化閥值搜尋，找出最佳閥值。對於晶片上既有雜訊，以區塊填補的演算法，將之填補、濾除，至此完成前背景的分離。接續將該影像進行水平與垂直投影，建立出網格，界定出每個雜交點可能出現的區域範圍。以Sobel運算子建出邊緣影像後，即可開始進行圓形霍氏轉換。圓的認定主要根據該影像所對應參數平面上的像素累積值，由於所需分析的影像複雜度較高，如何減少多餘的累計值，避免認定上的誤差是相當重要的。針對這部分，利用邊緣影像中梯度角的資訊，僅於參數平面上畫出與梯度角夾±90度的半圓。依此方法，不僅降低雜訊的干擾，判定圓所在的位置也更精準，也有效減少一半的運算時間。辨識出雜交點後，提供前、背景的像素強度資訊，便於微陣列資料的後續統計分析。將本研究建立之SPOTCapturer軟體的辨識結果與微陣列晶片影像常用分析軟體GenePix Pro 6.0比較之，SPOTCapturer準確率98.6%，召回率98.3%；GenePix Pro 6.0準確率97.5%，召回率97.9%，兩者以卡方檢定中之齊一性檢定驗證，證明確實存在顯著差異。對於雜交點的定位及尺寸估測方面，SPOTCapturer的準確度也較GenePix Pro 6.0高2.4%。取得訊號數值方面，考量到甜甜圈點造成的效應，加入前、背景間所應有的關聯性，除去範圍中的雜訊，以求獲得更精確之數值。綜合各項比較，透過SPOTCapturer，無論是在點的辨識或是值的擷取，的確可以做出更有效的分析，且不需給予分析影像繁複的晶片參數設定，提供了微陣列影像一個便利的分析平台。

In this research, we used circular Hough transform to be the core method to search circular spots in microarray images. Due to high sensitivity of circular Hough transform to noises in an image, noise removal plays a very important role in image pre-processing. At first, we used histogram equalization to enhance the contrast of images before image thresholding. As a result, it is easier to separate foreground pixels from the background using the Otsu’s method to find the best threshold value. Following image binarization, we used blob algorithm to filter noise pixels. Then we developed grids by vertical and horizontal histogram projections to determine the possible area of every spot in an image. The binary image was further processed with the Sobel operator to obtain the edge image that was further processed with the circular Hough transform to determine the position and boundary of each spot. The determination of circular spots was based on the number of pixels in the accumulating cells in the parameter space. To reduce computation complexity, it is important to avoid identification errors by decreasing redundant accumulations. Using the gradient angle information in an edge image, we were able to use only half circle, +90 and -90 degrees between gradient angles, in the circular Hough transform. With this approach, we can not only reduce the influences of noises but also improve the accuracy of spot identification. The computation time was also reduced in half. After identifying the spots, the intensities of foreground and background pixels were used in the subsequent statistical analysis of microarray data. Comparing the performance of SPOTCapture software developed in this research with the commercialized software GenePix Pro 6.0, SPOTCapture has the precision rate of 98.6% and the recall rate of 98.3% while the precision rate and recall rate of GenePix Pro 6.0 was 97.5% and 97.9, respectively. The performance difference between these two approaches was statistically significant as the results were tested with the chi-square test. As for the determination of position and area of spots in a microarray image, SPOTCapture has a higher accuracy than the GenePix Pro 6.0 by 2.4%. Our method is also sensitive in resolving the problems of donut spots frequently occurred in microarray images. In addition, the parameter settings of our method are less complicated and require less manual intervention. Thus, with all these advantages, the SPOTCapture can be used as an efficient platform for the analyses of microarray images.