Research On Recognition Algorithm To Evaluate Gene Amplification Status In FISH Image

Fluorescence In Situ Hybridization (FISH) is a techniques to detect cell gene’s amplification status. It has high stability, high accuracy and sensitivity advantages. In recent years, FISH is widely used to detect the amplification status of HER2 gene in tumor cells. The images strained by FISH techniques contain three parts:cell area, outside of cell area and red or green signal points. Currently, pathologist use microscope to count the number of red and green signal points. Then they judge whether HER2 gene is amplified by the ratio of red and green ones. The judgment method based on artificial, so it is vulnerable to the humor factors and its detection process is more tedious. It has important social significance and high application value to research automatic identification of FISH cells and the automatic cancer judgment.Many FISH cells are adhered and overlapped, the boundaries of them mainly fuzzy. To overcome the difficulty caused by the overlapping of cells in the image analysis for cancer diagnosis, this paper presents a new edge detection of tumor FISH cells which based on automatic random walk algorithm to separate the clustering cells. Firstly, detect cell and signal regions by the automatic threshold method which based on the RGB color model. Then, the regions of the cell’s seeds are detected by the ultra-erosion with some priori knowledge. Finally, the effective pixel markers are extracted from the detected regions and took as the object seeds in the random walk algorithm. The experiment shows that the method can detect cells well and segment multiple cells in a random walk way.The gene amplification status of FISH images is closely related to the number of the red and green signal points. So the signal points’classification and recognition is very important. This paper firstly extract signal points’ feature, then presents a new Laplacian Eigenmaps method(S-LE) on the foundation of traditional LE algorithm which can self-regulate it’s neighborhood parameter. The S-LE algorithm determines neighborhood size by manifold curvature to overcome the problem that traditional Laplacian Eigenmaps(LE) algorithm is not suitable for the distributed and non-smooth data. The experiments show that the algorithm is effective. This paper uses the S-LE algorithm to reduce the dimension of the morphological characteristics which extracted from the signal points. And then put the features to BP neural network to classify the signal points. The method not only saves time but also improves the recognition rate of the classifier.The judgment of tumor severity is the final step of the image detection. It is important to research how to get correct results by the number of particles of the red and green signals point. The judgment method based on global Ratio value and is most used currently, but it has some deficiencies and only suitable for artificial. So the paper designs a method base on cells’ Ratio value. It first calculates the Ratio value of the red signal points and the green ones in every cell. Then divide cells into normal and abnormal ones. Finally, evaluate the gene amplification status by the number of the category cells. The comparative experiments show that the method based on cell Ratio could obtain better results.