A Computer Aided Diagnosis Technique Of Retinopathy In Retinal OCT Images Under Limited Supervision

Retinopathies represent a major threat to human vision.Early diagnosis and treatment are the most effective ways to prevent vision deterioration.Optical coherence tomography(OCT)can perform three-dimensional imaging of the retina and has been widely adopted in clinical observation.Nevertheless,a main characteristic of OCT is its high data redundancy.For example,an OCT system with an axial resolution of 15μm acquires 512 two-dimensional tomograms from an 8 mm×8 mm retinal area.Consequently,indicating lesions in these images is challenging,even for specialists.Moreover,manual reading of retinal images requires ample clinical experience and professional knowledge because the size,shape,and location of lesions are diverse and complex.Recently,deep-learning-based computer-aided diagnosis has achieved outstanding performance,and it is being gradually adopted in clinical practice to assist specialists for fast and accurate diagnosis of diseases.However,most deep learning methods require supervised learning.Thus,their performance depends on the number and details of lesion labels,and the generality of the methods may be limited by data scarcity,thus preventing their use in clinical practice.We conducted studies on image quality assessment,retinopathy detection,and lesion quantification in OCT considering limited supervision.The studies are summarized as follows:(1)Retinopathy screening based on transfer learning:No automatic quality assessment method for retinal OCT images is available based on deep learning,and existing disease classification methods only focus on specific retinopathies.Hence,we propose a deep-learning-based primary retinopathy screening method that comprises stages of OCT image quality assessment and retinopathy detection.To balance the data volume and ability of the method,we adopt transfer learning to train the two stages.Using statistical analysis,quality assessment classifies three quality characteristics,and disease detection distinguishes pathological retinas from healthy retinas to increase generality.Experimental results demonstrate that the disease detection stage recognizes various retinopathies with an accuracy of 93.75%,and the detection accuracy is further improved by 3.75%on the dataset subject to quality assessment,indicating the importance of this stage.(2)Retinal lesion segmentation based on weakly supervised learning:As the segmentation map of retinal lesions is difficult to obtain and existing supervised methods only detect learned lesions,we propose a weakly supervised learning method for retinal lesion segmentation.The method provides pixel-level lesion segmentation from image-level labels for training and can recognize various types of lesions.We use coarse-grained image-level labels to train a generative model that can reconstruct a healthy retina from a pathological one in whole-slide OCT images.Then,lesions can be segmented by subtracting the original pathological retina from the reconstructed healthy retina.In principle,the proposed method can detect all abnormal areas,being promising for identifying unusual lesions.In addition,weak supervision substantially reduces the cost of labeling.We use a CycleGAN as the baseline and introduce a UNet-based generator and a dilated-convolution-based discriminator to perform reconstruction of pathological retinas.In addition,a structural similarity-based loss function contributes to the generation of realistic images.Experimental results demonstrate that the proposed method can reconstruct a whole-slide OCT image in 0.039 s,being much faster than existing methods,and its lesion segmentation is highly consistent with manual segmentation,providing a Dice coefficient of 0.8239.(3)Retinopathy detection based on unsupervised domain adaptation:Deep learning methods are sensitive to signal distribution and may perform poorly for outof-distribution test samples.Moreover,training for unseen test data requires labeling,which is labor-intensive and costly.We propose an adversarial-training-based approach for unsupervised domain adaptation to achieve pathological retina detection in crossdomain OCT images.We combine the cross-entropy loss and Wasserstein distance to obtain the domain distance,thereby stabilizing training and improving the detection performance.By integrating domain adaptation into retinopathy classification,the method trained with adversarial learning extracts domain-invariant and categorical features.Experimental results show that the classification accuracy on test data reaches 95.53%after domain adaptation.(4)Lesion detection based on unsupervised domain adaptation:Some retinal lesions cannot be quantified by a segmentation method and domain shift in lesion detection.Hence,we propose a method for retinal lesion detection based on unsupervised domain adaptation.The method can recognize and locate multiple lesions,such as choroidal neovascularization and retinal atrophy,in cross-domain OCT images.Faster-RCNN is the baseline for lesion detection,and domain adaptation is implemented for extracting global and instance-level features.A Wasserstein critic and domain classifier are used to perform domain adaptation.Experimental results show that the lesion detection accuracy on test data reaches 76.01%after domain adaptation,corresponding to an improvement by 8.31%compared with the baseline.In conclusion,we consider machine learning using limited supervision for quality assessment,retinopathy detection,and lesion quantification in OCT images,aiming to achieve accurate computer-aided diagnosis with limited data and labels available.This study can pave the way for the clinical application of deep-learning-based computeraided diagnosis of retinopathies.