Research On Automatic Diagnosis Of Autoimmune Diseases And Fundus Diseases

With the continuous development of artificial intelligence technology,many industries have hatched applications related to artificial intelligence.The artificial intelligence & medical treatment is the one of the most important research directions.The application prospect of artificial intelligence in medicine and health field is attracting more and more attention.At the same time,many applications of artificial intelligence technology are developed,which carries out specific tasks with medical imaging data in medical field,such as early screening,auxiliary clinical diagnosis,etc.As a representative technology of artificial intelligence,deep learning has been widely used in academia and industry,which is a key technology breakthrough and devoted to specific practical application.In medicine field,it is a very significant work to use the related technology of artificial intelligence to implement the early screening of diseases.It can not only reduce the workload of medical staff,but also achieve rapid and portable examination and diagnosis.Meanwhile,with early screening,doctors can also timely intervene and treat the disease in the early stage of disease development.It has become an irreversible trend to develop an automatic diagnosis system by leveraging the related technology of artificial intelligence to assist clinicians to make scientific medical diagnosis,which can save medical resources and effectively address the problem of missed diagnosis and misdiagnosis caused by doctors' subjective factors.This paper focuses on the algorithm and application of deep learning in the fields of autoimmune disease detection and fundus disease screening and detection,aiming to develop an automatic diagnosis model for autoimmune disease detection using HEp-2 cells and an automatic diagnosis model for fundus disease screening and diagnosis using scanning laser ophthalmoscope(SLO)images.The main work and contributions consists of the following aspects:(1)Deeply Supervised Full Convolution Network(DSFCN)for HEp-2 Specimen Image SegmentationA DSFCN framework for automatic segmentation of HEp-2 sample image is proposed,which can provide accurate target area information for subsequent HEp-2 cell classification task.The framework utilizes the VGG-16 model,which is pre-trained on PASCAL VOC 2012 dataset,and makes full use of dense deconvolution and multilayer supervision module.Dense deconvolution layer can enable the network to learn a lot of shallow feature information and reuse it,so that the decoder module can preserve the edge and contour information while recovering the resolution,which further improves the segmentation performance.With the hierarchical supervision module,the network can effectively learn the correlation between the pixel prediction value and the ground-truth so that the segmentation results are further refined.Extensive experiments demonstrate that the proposed DSFCN segmentation framework can segment the region of interest very well,and has a promising segmentation performance for the segmentation task of HEp-2 specimen images.(2)Joint Segmentation and Classification Task via Adversarial Network for HEp-2 Cell ImagesThe purpose of analyzing the staining patterns of HEp-2 cells is to carry out the automatic diagnosis of autoimmune diseases.The accurate segmentation results of HEp-2 cells can provide rich and effective feature information for the classification of Hep-2 cell staining patterns.Therefore,guided by the tasks,a novel hybrid framework for HEp-2 cell image segmentation and classification is proposed.It contains three modules,the segmenter module,the discriminator module,and the classifier module.The first two and third modules form the GANs for segmentation and classification tasks,respectively.For the segmenter,the ResNet-34 model being pre-trained on ImageNet dataset is employed as the encoder to extract the multi-scale information and the MS-ASPP is also used to refine the spatial information to obtain rich boundary information so that the segmentation performance can be improved,which provides better service for classification task.For the discriminator and classifier,the modified MobileNetv3 that is built by increasing the number of channels in the middle hidden output layer of the original mobilenetv3 model and changing the size of the convolution kernel,which is called ACM-Net,which realizes discrimination and classification tasks at the same time.In particular,the features from MS-ASPP are used as the auxiliary classifiers to supervise the main classification networks to obtain better classification performance,further achieve more accurate automatic diagnosis of autoimmune diseases.(3)Attention Encoder and Multi-branch Structure Based Generative Adversarial Networks for Fundus Disease Detection using SLO ImagesThrough the research of the previous work,we observe that the GANs has a strong operability in the field of feature extraction and image synthesis.Hence,in order to solve the problem of small amount of data in SLO image and realize the detection task of abnormal fundus image,guided by the method and thought of the previous work,a novel generative adversarial network called AMD-GAN is proposed to detect the diseased fundus images by classifying the abnormal and normal SLO images.Specifically,the designed generator of AMD-GAN consists of two parts: the attention encoder(AE)and generation flow network with residual up-sampling(RU)block.With the AE module,the generator can produce the fake features with the same size as the features from real images,which are fused with the attention module.As for discriminator,we copy the route of the last two layers of feature extraction of the original ResNet-34 model to comprise another branch to build the multi-branch(MB)ResNet-34 framework,which makes the proposed network can extract more high-level feature information than the traditional ResNet-34 model,and further boost the detection performance.Furthermore,the depth-wise asymmetric dilated convolution(DADC)is utilized to extract local contextual features and accelerate the training processing.Meanwhile,the last layer of discriminator is modified to build the classifier to recognize the diseased and normal SLO images.The extensive experimental results demonstrate that our method is effective for the detection of diseased and normal images of the SLO datasets,which achieves the best detection performance by exploring the experts' knowledge.(4)Cross-attention Multi-branch Network for Fundus Diseases Classification Using SLO ImagesThe proposed AMD-GAN framework can effectively distinguish whether the fundus SLO image is a normal fundus image,which realizes the primary screening of abnormal fundus disease.However,clinically,in order to realize the automatic diagnosis of fundus diseases,more attention will be paid to the specific fundus diseases of patients.For this purpose,a novel cross-attention multi-branch network based on ResNet-34 model is proposed to accomplish the classification task of the selected three diseases(i.e.Retinitis pigmentosa(RP),Diabetic retinopathy(DR)and Coats)and normal images using SLO fundus images.The devised framework mainly includes four parts: multi-branch network,Atrous spatial pyramid pooling(ASPP),depth-wise attention and cross-attention modules.Specifically,the multi-branch employs the ResNet-34 model as the backbone to expand the network branch to extract more deeplevel features.Following the last feature extraction layer of the first two branch networks,the ASPP is utilized to extract the multi-scale features via setting different dilation rates to gain the rich spatial contextual feature information.The depth-wise attention module is devised to fuse the features from the ASPP module and another branch network with original ResNet-34 model so that the whole framework can preserve the global information while focusing on the target areas of interest.To make full use of the extracted features from the multi-branch network and enforce the network to highlight the characteristics of the specific disease and suppress the irrelevant areas,we design a cross-attention module to fuse the channel attention maps and spatial attention maps from the first two branch networks through the crisscross connection manner.