The Design Of Classifier On Gastric Mucosa Tumor Microscopic Image

With the wide application of computer, it's an important topic in the field of medicine image to make research on tumor microscopic image auto-recognition with computer pattern recognition technology. As the inherent complexity of medical images, the severe adhesion between cells, glands, cells and glands, the classification performance of a single classifier based on a single feature is often difficult to meet the clinical requirements. In order to improve the performance of single classifier, this paper designs the classification system based multi-feature fusion and multi-classifier fusion for gastric mucosa tumor microscopic images.Feature extraction is the precondition of classifier design. According to the diagnostic basis provided by pathologist, by analyzing the various features of images, we extract four gland features, seven cell features and the image pixel color value in the R channel, fuse these features, as an basis of the classifier design.Gastric microscopic images are divided into three categories: normal, cancerous and dysplasia. In order to classify them automatically, we design a two-level classifier: the first-level classifier classify the images into normal or abnormal (including cancerous and dysplasia); the second-level classifier classify the abnormal images obtained from the first-level classification into cancerous or dysplasia.The first-level classifier is the fusion of the global classifier based on gland features and the local classifier based on cell features. The global classifier also is the fusion of one Mahalanobis distance classifier based on for gland features and two minimum distance classifiers based on single gland feature, the three classifiers is progressive arrangement. The local classifier is a decision tree constructed by the cell features. Firstly, we classify the image into normal or abnormal by the global classifier, if judged as normal, then, classify the image into normal or abnormal through the local classifier. We classify the abnormal image judged by the global and local classifier into cancerous or dysplasia.The second-level classifier is a pre-classifier based on the improved PCA+LDA, which is based on the image pixel color value in the R channel. In this method, Firstly, we use PCA for feature dimension reduction, the feature values are projected on to the PCA sub-space. In the PCA sub-space, firstly, we suppose that the test image belongs to cancerous, then, do LDA transform for the test image and the training samples at same time, in the LDA sub-space, calculate the distance between the test image and the mean of the cancerous training samples, namely d1; Then, we assume that the test image belongs to dysplasia, do the same operation like that, calculate the distance between the test image and the mean of the dysplasia training samples, namely d2; At last, compare the distance d1 and d2, namely, compare the within-class compact ratios in different assumptions, the smaller the distance is, the higher the within-class compact ratio is, so the test sample were judged to that class. The improved PCA+LDA well overcome the shortcomings that traditional PCA+LDA is poor generalization ability for test sample.We do the test for the classification system, get a high classification accuracy rate, it has a practical significance for medical research and clinical diagnosis and a good application prospects.