Research On Anomaly Detection Of OCT Image Based On Epistemic Uncertainty

As a non-contact and high-sensitivity imaging technology,optical coherence tomography（OCT）technology is widely used in the detection of retinal diseases.The research on computer aided diagnosis technology for retinal diseases has increasingly become a research hotspot in related fields.With the development of deep learning technology,the detection method based on convolutional neural network（CNN）has realized the efficient classification of common retinal diseases.However,these methods rely on large-scale OCT images with retinopathy as training set,for diseases with low incidence,these methods are difficult to achieve effective diagnosis of retinal diseases with low incidence.Most of the existing detection methods for non-specific fundus diseases are based on the generative adversarial networks（GAN）.However,the accuracy and sensitivity of these methods do not reach a high level.This paper proposes an anomaly detection method,named EMADM,for retinal diseases in OCT images based on epistemic uncertainty.The basic principle of the anomaly detection method is to train a network model for learning the features of normal OCT images.When the OCT image with lesions is input as test data,the difference between the output and the output of the normal image is detected,and the difference is used as the basis for judging whether the input data is abnormal.In the EMADM method,the difference between the lesion and the normal tissue area is quantified by the uncertainty prediction result obtained from the Bayesian neural network.Bayesian neural networks can quantify the uncertainty of the model predictions.Therefore,this paper uses the epistemic uncertainty results obtained from the Bayesian neural networks as the standard for anomaly detection in the EMADM method.The basic process of the method is as follows: First,a Bayesian neural network model MBUNET is trained using the retinal OCT images from healthy people as the training set to extract the features of normal retinal OCT images.Once an OCT image with lesions is input,since the lesion features are not learned and extracted in the training stage,the lesion area lead to the high uncertainty.Then,the influence of the aleatoric uncertainty generated in some normal tissue regions is eliminated by the Borderline uncertainty filtration and uncertainty image segmentation methods in this paper.Finally,we design a threshold-based anomaly detection function.If the uncertainty result of the input data is higher than the threshold,it is predicted as abnormal data,otherwise it is predicted to be normal.In this paper,the effectiveness of the EMADM method is verified by the comparative experiments.The experimental results show that the EMADM method is superior to the existing anomaly detection algorithms in the accuracy,sensitivity and specificity.At the same time,other relevant experiments have proved that each step in the anomaly detection process contributes to the improvement of the performance of the method,and verifies the effectiveness and necessity of each step.