Research On Retinal OCT Image Processing Method Based On Semi-Supervised Learning

Optical Coherence Tomography(OCT),a non-invasive and high resolution biological tissue imaging technique,has been widely adopted in the retinal diseases’ diagnosis and evaluation field.However,due to the inherent characteristics of retinal OCT imaging technology,retinal OCT images often tends to contain a large amount of speckle noise,leading to the decline of imaging quality and affecting the subsequent image processing.Furthermore,the small proportion of retinal regions in OCT images and the complexity of pathological characteristics,are also great challenges to research on retinal OCT image processing method.Besides,the lack of pixel-level labeled data which are time-consuming and expensive,has always been an important factor limiting the performance improvement of deep learning based methods.How to further improve the performance of retinal OCT image processing methods with unlabeled data under the condition of limited labeled data has become one of the research hot topic.Therefore,focusing on the method research of image denoising enhancement and image segmentation in retinal OCT image processing by combining semi-supervised learning method.The main research contents and innovations are summarized as follows:(1)A novel semi-supervised capsule condition generation adversarial network(Semi Caps-cGAN)is proposed for speckle noise suppression and contrast enhancement in retinal OCT images.In order to improve the model’s ability to capture complex structural features in retinal OCT images,and avoid overfitting problems caused by the sudden increase in the number of parameters,a novel capsule conditional generative adversarial network(CapscGAN)with less parameters is designed as the main body network of Semi Caps-cGAN;To handle the problem of the loss of retinal structure information in OCT images due to the lack of corresponding ground truth guidance for unlabeled data during the training process,a novel joint semi-supervised loss function is developed to optimize the proposed method,helping model suppress speckle noise and focus on the retina structure characteristic of OCT images of retina.Finally,comprehensive experimental results show that the proposed method achieves better performance than other excellent denoising methods,and can be well applied to images obtained from different OCT scanners and acquisition modes,which proves the effectiveness and robustness of the proposed method.(2)Focusing on the problems and challenges in the joint segmentation task of macular hole(MH)and cystoid macular edema(CME),such as complex pathological features,blurred boundary,and the insufficient of data with pixel-level annotation ground truth,a novel self-guided optimization semi-supervised segmentation method(Semi-SGO)is proposed.First,aiming at complex pathological features of MH and CME in retinal OCT images,a novel dual decoder dual task fully convolutional network architecture(D3T-FCN)is designed based on the principle of auto-encoder and combining the structural characteristics of U-Net.The D3T-FCN can improve the network’s ability to extract complex pathological features through the dual-task joint optimization of feature reconstruction and lesion segmentation,thereby improving the accuracy of joint segmentation of MH and CME.Then,based on the designed D3T-FCN and combined with the knowledge distillation teacher and student strategy,a self-guided optimization semisupervised segmentation method named as Semi-SGO is designed to further improve the segmentation accuracy of MH and CME in retinal OCT images by leveraging unlabeled data.Finally,comprehensive experiments are conducted to evaluate the performance of the proposed method.Compared with the state-of-the-art deep learning based methods,the proposed Semi-SGO can significantly improve the accuracy of MH and CME in retinal OCT images.(3)Aiming at the problems of small lesion size,large distribution difference,diversified morphology and lack of pixel-level annotated data in the drusen segmentation task of retinal OCT images,a novel semi-supervised multi-scale transformer global attention network(Semi-MsTGANet)based on the pseudo-lable augmentation semisupervised strategy is proposed to segment drusen in retinal OCT images.Firstly,to tackle the complex pathological manifestations problem of drusen in retinal OCT images,a novel multi-scale transformer global attention network named as MsTGANet is designed to improve the drusen segmentation accuracy in retinal OCT images.In MsTGANet,two newly proposed modules of multi-scale transformer non-local(MsTNL)and multisemantic global channel and spatial joint attention(MsGCS)are adopted to improve the model’s ability to learn multi-scale non-local features with long-dependent information and to obtain multi-semantic global context features.Then,aiming at the problems of insufficient pixel-level annotated data and the waste of a large amount of unlabeled data,a novel semi-supervised segmentation method of Semi-MsTGANet is proposed based on the newly proposed MsTGANet and combined with pseudo-label augmentation semisupervised strategy,which can leverage a large amount of unlabeled data to further improve the drusen segmentation accuracy in retinal OCT images.Finally,the experimental results show that compared with other excellent segmentation methods,the drusen segmentation accuracy of Semi-MsTGANet has been significantly improved,which proves the effectiveness of the proposed method.(4)In order to further evaluate the generality and clinical significance of the proposed methods,the performance of the proposed Semi-SGO and Semi-MSTGANet in different retinal OCT image segmentation tasks are first further verified by experiments,and compared with the commonly used semi-supervised segmentation frameworks.The experimental results show that both the proposed Semi-SGO and Semi-MSTGANet methods obtains better performance in different retinal OCT image segmentation tasks,among which the Semi-MSTGANet achieves the best comprehensive performance in the jomt segmentation of MH and CME in retinal OCT images and the segmentation of drusen,demonstrating the effectiveness and generality of the proposed methods.Furthermore,by the cooperation with clinical ophthalmologists,an automatic analysis of clinical indicators such as lesion diameter,height,and volume was carried out based on the segmentation results of MH and CME segmented by Semi-MsGANet.The experimental results show that there is good consistency between the clinical indicators measured based on the SemiMsTGANet segmentation results and the clinical indicators manually annotated by the ophthalmologists,demonstrating that the proposed semi-supervised retinal OCT image segmentation method can provide important reference for ophthalmologists in evaluating retinal diseases.