A multi-scale anomaly detection framework for retinal OCT images based on the Bayesian neural network. (May 2022)

• Record Type: Journal Article
• Authors: Mou, Lintao; Liang, Lingling; Gao, Zhanheng; Wang, Xin
• Is Part Of: Biomedical signal processing and control. Volume 75(2022)
• Publication Date: 2022-05
• Subjects: Bayesian neural network -- Deep learning -- Medical image analysis -- Optical coherence tomography -- Epistemic uncertainty; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.103619
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: A Bayesian neural network architecture for anomaly detection. Multi-scale Monte Carlo sampling improves the accuracy of uncertainty estimation. Borderline uncertainty filtration reduces the uncertainty from healthy regions. Detecting lesions after segmenting the uncertainty graph improves the sensitivity. Abstract: As a non-contact imaging technology, optical coherence tomography (OCT) is widely used in retinal disease detection. There is a growing interest among researchers for automated detection of retinal diseases using OCT images. With the development of deep learning, convolutional neural network (CNN) enables accurate detection of common retinal diseases. However, large-scale labeled data is necessary for these deep learning methods. It limits the application of these methods in retinal disease detection, especially for diseases with low occurrence. Therefore, we propose an automated anomaly detection method for unspecific retinal disease, and only healthy OCT images are needed for training. The method uses epistemic uncertainty as the criterion of anomaly detection. Our method assumes that the lesion correlates with a high epistemic uncertainty region. We propose a Bayesian neural network (BNN) model, Multi-scale Bayesian U-Net (MBU-Net), to obtain epistemic uncertainty of OCT images. We then design an algorithm to reduce the uncertainty generated by healthy tissue regions. Finally, a threshold-based function is designed to distinguish whether the input data is healthy or not. We evaluate our approach on a public dataset, UCSD dataset, and a private dataset collected by ourselves, named BFHJLU dataset. The experimental results show that the proposed method can achieve 94.9% accuracy, 97.9% sensitivity, and 86.0% specificity on the UCSD dataset. On the BFHJLU dataset, the accuracy, sensitivity, and specificity are 92.1%, 97.3%, and 87.7%. The proposed method outperforms existing anomaly detection methods.