深度学习在眼底图像分析中的研究与应用

RESEARCH AND APPLICATION OF DEEP LEARNING ON FUNDUS IMAGE ANALYSIS

王思懋

11749124

硕士

信息与通信工程

唐晓颖

2020-05-29

2020-07-24

哈尔滨工业大学

深圳

视网膜血管网络是唯一可以在体内可视化和拍照的血管系统。视网膜血管成像能够为患有特定心血管疾病和眼科疾病的患者提供临床预后信息。分割视网膜血管是监测视网膜血管网络状况的先决条件。青光眼是一种广泛的眼病，发病率高，导致视力下降。由于这种疾病的恶化是不可逆的，因此及早和及时的诊断非常重要。杯盘比（CDR）是用于诊断青光眼的最常见生物标志物之一。因此准确的分割视杯视盘及眼底血管并对其进行形态学分析具有重要的应用价值和实际意义。在分割眼底图像之前，本文首先对原始眼底图像进行预处理。包括图像的通道选择、滤波去噪、血管质量评估等。预处理之后，采用深度学习网络实现了眼底图像的血管分割、视盘分割、视杯分割。分割完成之后再对各个结构进行形态学分析。具体研究内容如下：（1）提出并验证了一种无监督的血管质量评估方案，用于预处理中的图片筛选。在进行分割任务之前，可以有效剔除对算法精确度产生不良影响的图片，同时可以为某些疾病诊断提供诊断依据。（2）提出了一种用于OD和OC联合分割的新颖框架。我们工作的主要贡献在于：为了消除不确定性，我们从传统贝叶斯神经网络（BNN）的最大似然估计（MLE）中学习，并采用新颖的框架（包括分段网络和不确定性）来实现估算网络。结合迁移学习的训练策略，在加快网络收敛速度的同时提高分割精度。（3）将上述网络应用于血管分割中。血管分割中利用可分离的空间和通道流以及密集的相邻血管预测来捕获血管之间的最大空间相关性。在视盘和视杯分割中，在训练和预测阶段都使用了Geometric transformations和Overlapped patches，以有效地利用在训练阶段学到的信息并细化分割。（4）针对上述分割结果提出了一种形态学分析方法。针对血管分割结果可以精确计算血管的直径、长度、密度等，其中最为重要的是血管直径的计算。我们提出了一种基于血管中心线的血管直径测量方法。针对视杯视盘的分割结果可以精确计算杯盘比（视杯与视盘的直径之比）等数据。最后，本文在三个公开数据库DRIVE、REVIEW和ORIGA上，对上述研究内容进行了实验验证，并与现有研究结果进行了比较。结果表明，本文中所提出的血管分割及视杯视盘分割方法具有较高准确度和稳定性，而对于血管和视杯视盘的形态学分析具有较高的精确度。

The retinal vascular network is the only vascular system that can be visualized and photographed in vivo. Retinal vascular imaging can provide clinical prognostic information for patients with specific cardiovascular and ophthalmic diseases. Segmentation of retinal blood vessels is a prerequisite for monitoring the condition of the retinal blood vessel network. Glaucoma is a widespread eye disease with a high incidence, which leads to decreased vision. Because the deterioration of this disease is irreversible, early and timely diagnosis is very important. Cup to disc ratio (CDR) is one of the most common biomarkers used to diagnose glaucoma. Therefore, accurate segmentation of the optic disc optic disc and fundus vessels and their morphological analysis have important application value and practical significance. Before segmenting the fundus image, this paper first preprocesses the original fundus image. Including image channel selection, filter denoising, blood vessel quality assessment, etc. After pre-processing, deep learning network was used to realize blood vessel segmentation, optic disc segmentation and optic cup segmentation of fundus image. After the segmentation is completed, the morphological analysis of each structure is performed. The specific research contents are as follows:(1) We have proposed and verified an unsupervised vascular quality assessment scheme for image screening in pretreatment. Before the segmentation task, it can effectively remove the pictures that have an adverse effect on the accuracy of the algorithm, and at the same time can provide a diagnosis basis for the diagnosis of certain diseases.(2) A novel framework for joint segmentation of OD and OC is proposed. The main contribution of our work is: In order to eliminate uncertainty, we learn from the maximum likelihood estimation (MLE) of traditional Bayesian neural networks (BNN), and adopt novel frameworks (including segmented networks and uncertainties) To realize the estimation network. Combined with the training strategy of transfer learning, the segmentation accuracy is improved while accelerating the network convergence speed.(3) Apply the above network to blood vessel segmentation. Vessel segmentation utilizes separable space and channel flow and dense adjacent vessel prediction to capture the maximum spatial correlation between vessels. In the segmentation of optic disc and optic cup, both geometric transformations and Overlapped patches are used in the training and prediction stages to effectively use the information learned in the training stage and refine the segmentation.(4) A morphological analysis method is proposed based on the above segmentation results. According to the blood vessel segmentation results, the diameter, length, density, etc. of the blood vessel can be accurately calculated, the most important of which is the calculation of the blood vessel diameter. We propose a method for measuring the diameter of blood vessels based on the center line of the blood vessels. For the segmentation results of the optic disc optic disc, data such as the ratio of the optic disc to the optic disc (the ratio of the diameter of the optic cup to the optic disc) can be accurately calculated.Finally, in this paper, on three public databases DRIVE, REVIEW and ORIGA, the above research contents were tested and compared with the existing research results. The results show that the blood vessel segmentation and optic disc segmentation methods proposed in this paper have high accuracy and stability, while the morphological analysis of blood vessels and optic disc optic discs have high accuracy.

深度学习 血管分割 视盘分割 眼底图像 视杯分割 形态学分析

Deep learning blood vessel segmentation optic disc segmentation fundus image optic cup segmentation morphological analys

中文

联合培养

南方科技大学

王思懋. 深度学习在眼底图像分析中的研究与应用[D]. 深圳. 哈尔滨工业大学,2020.