A Novel Algorithm Of Sub-cellular Localization Based On Fluorescence Microscopy Images

Protein subcellular localization is an important reseacrh subject of molecular cellbiology arid proteomics. In order to play its biological function, protein must betransported to specific organs after compounded at the ribosome, and then the wholelife of the body could be in working order. The functions of proteins are related withtheir subcellular localization. Thus, we can get the help of inferring the performanceand function of the protein by knowing the location information of subcellular.Automated fluorescence microscopy is one of the most beneficial means forspeciifc localization of proteins and other biological macromolecules qualitativeresearch at light microscopy level, and it allow protein localization imaging inhigh-throughput. Therefore, we need an automated and efifcient way to effectivelyquantify, distinguish and classify subcellular images, the key point is that imagefeature extraction and classification algorithm designing.This paper mainly focus on the two themes, and have a research on image featureextraction and algorithm of classification. Finally, we tested and analyzed on differentdata sets. The main innovation of this paper are as follows:This paper presents a novel protein subcellular image extraction algorithm basedon local invariant features, which has character of rotation and translation in vairance.The main idea is: first，a threshold is selected for image pre-processing, these pixels,whose value are less than the threshold, are setting to zero, those pixel whose value isgreater than or equal to the threshold remains unchanged，we divided the image tobackground and target portions, We try not to loss image information, so that it canreduce workload of feature extraction and improve the efficiency of algorithm;Second, to each pixel in the target area, the differences between its grey value and thegray value of surrounding eight pixels in the neighborhood are calculated, respectively.Then we marked1to these surrounding pixels whose values are equal to or greaterthan the central、and0are to those whose values are less. We will get a adjacentbinary model, and the operation of the summing for the model is obtained(ie, the number of1), and its value ranges from0to8，so that each image pixel can begrouped into nine groups of0-8; eventually, wc calculate the number of each groupfor each image, and then calculate the positive, negative and absolute difference basedon the adjacent model. So that we can get four statistics rfom each group and9x4=36statistics are obtained for each image.In this article, we developed the support vector machine as a forecasting tool,tested on the three different image data sets called the LOCATE Endogenous,LOCATE Transfected and2D HeLa. The classification accuracy was96.7%,92.3%and90.2%, respectively. The result shows, the method of this paper has a distinctadvantage and made a better prediction results compared with other methods.