Deep Learning Based Hep-2 Cell Image Analysis

Antinuclear Antibody(ANA)test is a common protocol for many Autoimmune diseases,e.g.systemic Lupus Erythematosus,Sjorgrens syndrome and rheumatoid arthritis etc.Human epithelial type 2(HEp-2)cells with Indirect Immunofluorescence(IIF)are the gold standard for the ANA test.In clinical diagnosis,at least three physicians are required to observe the HEp-2 specimen images of these patients by fluorescence microscopy and determine the stained patterns by voting.This approach is super time consuming when the amount of specimen images is large.At the meantime,due to the subjective experience of the physicians and changes in imaging conditions(such as staining methods,etc.),the test results may not be very objective.Conventionally,most HEp-2 cell image analysis systems are based on the hand-crafted features,which may lead to lower performance.Recently,deep convolutional neural network(CNN)based methods have achieved excellent results in many tasks such as natural image recognition,image generation,and object detection.Inspired by the success of CNNs,this thesis studies deep convolutional network based HEp-2 cell image analysis system.This thesis tackles three tasks,i.e.HEp-2 cell image classification,HEp-2 specimen image segmentation and HEp-2 specimen image classification.The contributions are summarized as follows:1.A HEp-2 cell image classification system based on Deep Cross Residual Network is proposed.We propose cross residual connection module based on the residual connection of the ResNet and successfully train the deep cross residual network with a class-balanced data augmentation method.State of the arts results are achieved on two benchmark datasets without using any external data.2.A HEp-2 specimen segmentation system based on two series generative adversarial networks(GANs)is proposed.The segmentation task from the fluorescently stained HEp-2 specimen image to the binary mask image is completed with an image translation way.The proposed system outperforms all existing GANs based approaches.3.The Mask R-CNN algorithm is used to segment and classify HEp-2 specimen images,and the efficient spatial pyramid module with dilated convolution and soft non-maximum suppression algorithm are introduced to improve the results of Mask R-CNN.