Segmentation of microarray images using pixel classification-comparison with clustering-based methods

DNA microarray technology yields expression profiles for thousands of genes, in a single hybridization experiment. The quantification of the expression level is performed using image analysis. In this paper we introduce a supervised method for the segmentation of microarray images using classification techniques. The method is able to characterize the pixels of the image as signal, background and artefact.

The proposed method includes five steps: (a) an automated gridding method which provides a cell of the image for each spot. (b) Three multichannel vector filters are employed to preprocess the raw image. (c) Features are extracted from each pixel of the image. (d) The dimension of the feature set is reduced. (e) Support vector machines are used for the classification of pixels as signal, background, artefacts. The proposed method is evaluated using both real images from the Stanford microarray database and simulated images generated by a microarray data simulator. The signal and the background pixels, which are responsible for the quantification of the expression levels, are efficiently detected.

A quality measure (qindex) and the pixel-by-pixel accuracy are used for the evaluation of the proposed method. The obtained qindex varies from 0.742 to 0.836. The obtained accuracy for the real images is about 98%, while the accuracies for the good, normal and bad quality simulated images are 96, 93 and 71%, respectively. The proposed classification method is compared to clustering-based techniques, which have been proposed for microarray image segmentation. This comparison shows that the classification-based method reports better results, improving the performance by up to 20%.

The proposed method can be used for segmentation of microarray images with high accuracy, indicating that segmentation can be improved using classification instead of clustering. The proposed method is supervised and it can only be used when training data are available.