基于深度学习的视网膜OCT图像无监督去噪方法

IF 1.6 4区物理与天体物理 Q3 OPTICS

光学学报 Pub Date : 2023-01-01 DOI:10.3788/aos230720

吴广义 Wu Guangyi, 袁卓群 Yuan Zhuoqun, 梁艳梅 Liang Yanmei

引用次数: 0

Abstract

以散斑噪声为主的噪声干扰严重影响视网膜光学相干层析（OCT）图像质量。深度学习是一种有效的去噪方法。但对活体成像而言，其很难获取多帧配准的真值图像，这影响了监督学习方法的效果。提出一种无监督深度残差稀疏注意力网络用于视网膜OCT图像去噪，并分别从视觉评价和数值评价两方面与传统的三维块匹配滤波去噪算法和经典的深度学习去噪网络进行对比。研究了监督学习与无监督学习策略下3种卷积神经网络的去噪性能，并利用公开的视网膜OCT图像数据集进行泛化能力测试。实验结果表明：所提算法的视觉评价和数值评价均具有良好的降噪效果，可以实现视网膜OCT图像高质量降噪，具有较强的泛化性，而且与监督学习相比，无监督学习在数据集不充分时仍能获得较好的降噪性能，可以有效地辅助医生进行准确高效的临床诊断。

查看原文本刊更多论文

基于深度学习的视网膜OCT图像无监督去噪方法

本文章由计算机程序翻译，如有差异，请以英文原文为准。

求助全文

约1分钟内获得全文求助全文

来源期刊

光学学报光学

CiteScore

2.80

自引率

37.50%

发文量

16537

期刊介绍： Researching is owned by Chinese Laser Press (CLP), which is established by Shanghai Institute of Optics and Fine Mechanics and Chinese Optical Society in 2009. Nowadays, CLP publishes 11 journals and manages three online platforms. Journal publishing activities include both traditional and digital models, and the first journal can be traced back to 1964. The CLP Online Library includes the CLP journals and partnered ones in China, and provides literature and intelligence services for users. The product platform, named as OEShow, connects sellers and buyers of optoelectronics products. Researching (formerly known as The CLP Publishing) is featured with stable operation and leading technology, collects CLP journals and partnered optics and photonics journals, and provides readers an optical publishing platform with global influence. CLP is on the way of building a modern publishing group combining traditional business and digital publishing.