Classification Method For HEP-2 Cells Based On Texture Features And Deep Learning

HEp-2 cel s with Indirect Immunofluorescence(IIF)are used as a substrate in human serum,which is a commonly-used technique for detecting anti-nuclear antibodies(ANA),and can be visualized via a fluorescence microscope.HEp-2 cell image classification plays an important role in the diagnosis of many autoimmune diseases.However,the traditiona l approach requires experienced experts to artificially identify cell patterns,which extremely increases the workload and suffer from the subjective opinion of physician.To address it,This paper studies the classification of HEp-2 cel s from two aspects of texture features and deep learning.First,study the classification of HEp-2 cell patterns based on texture features.Specifica lly,the pairwise rotation invariant co-occurrence local binary pattern(PRICoLBP),dense scale invariant feature transform(DSIFT)and color histogram are investigated to extract the cells characteristics.Differing from the traditional LBP-based feature,PRICoLBP not only captures spatial context information effectively,but also possesses rotational invariant characterist ics.DSIFT feature also has some desirable properties such as invariance to attacks.Then use Kmeans clustering method to reduce dimension,generate histogram statistical features,and finally use SVM for classification.The proposed method is validated on publicly availab le MIVIA HEp-2 cel s database.The method with the same evaluation method as ICPR2012-competition,cel-level accuracy is 70.90%,picture level accuracy is 84.71%.Second,the research is based on a deep residual network(ResNet)-based framework to automatically identify HEp-2 cells by migrating learning strategies.Compared with the existing methods utilizing either low-level hand-crafted features or DCNNs with shal ow learning networks,we adopt a residual network of 50 layers(ResNet-50)that are substantially deep to acquire rich and discriminative feature.The main structure of ResNet-50 is residual connection,which can solve the degradation problem effectively.Different from most of the deep learning models learnt from scratch,we use the transfer learning pre-trained from a very similar dataset(from ICPR2012 to ICPR2016-Task1)to fine-tune our own model.Our proposed framework achieves an average class accuracy of 92.63% on ICPR2012 HEp-2 dataset and a mean class accuracy of 94.87% on ICPR2016-Task1 HEp-2 dataset,which outperforms the traditiona l methods.The texture feature scheme proposed in the thesis compares the feature extraction scheme based on LBP extension and other literatures proposed improve the classification performance and effect.The deep learning scheme compares the HEp-2 cell classification of AlexNet,VGG-16 and VGG-19 networks.Optimize network training time and improve the classifica t io n performance of the target network model.