Automated Classification And Segmentation Method For HEp-2 Cell Based On Deep Level Feature Learning

Accurate Human Epithelial-2(HEp-2)cell image classification in many autoimmune diseases diagnosis and subsequent treatment plays an important role.For example,systemic rheumatic disease,multiple sclerosis,drug-induced lupus erythematosus,systemic lupus erythematosus and diabetes.Image segmentation is usually the first step in image classification and later diagnosis and treatment.The key challenge is the large intra-class and small inter-class and background noise caused by uneven lighting.In order to solve this problem,design a set of computer aided diagnosis system to help doctors automated segmentation,classification has become a hot topic,this article mainly for HEp-2 cell image automatic segmentation,classification algorithms,and proposes three different deep learning methods based on machine learning and deep learning.The first,we proposed the Pairwise Rotation Invariant co-occurrence Local Binary Pattern(PRICo LBP)method with strong classification ability and robustness.PRICo LBP not only has rotation invariance,but also can effectively capture spatial context symbiosis information.PRICo LBP by ensuring that the response of the symbiotic the coding strategy and effective rotation invariant in pairs of rotation invariance,encoding the symbiosis to reflect the local curvature information of the relative direction Angle ensures its powerful ability to distinguish and rotation invariance,and also by local gradient information by symbiotic mode with different weights,full use of the boundary and profile information.The second,we propose a Deeply Supervised Residual Network Framework(DSRN)to classify HEp-2 cell.Specifically,we used the residual 50 layer network(resnet-50)to basically extract the rich and differentiated features.The deep supervision mechanism can promote the training of the 50 layer network(resnet-50)directly under the direct guidance and the upper network to further improve the classification performance.At the same time we also joined the cross-modal transfer learning.A similar set of data sets is used to train and fine-tune our DSRN model to initialize our target network to speed up the convergence speed of the network,reduce computation and improve classification performance.The third,we put forward a framework of automatic cell image segmentation region,Dense Deconvolutional Network(DDN),it can solve the problem of localization and classification at the same time in this article.Our proposed models mainly include Dense deconvolution Layer(DDL),chain residual pool and hierarchical monitoring mechanism.DDN can learn the distinguishing features and effectively integrate multi-level contextual information.Our method can automatically segment the interest area of HEp-2 cell image.And then to extract the effective features for classification,this to no prior knowledge of a new image,to be able to accurately extract the feature of interest,so as to realize the ascension of classification performance is a deep scientific research value and the clinical value.