结合多尺度特征融合的扩张残差U-Net分割网络在视网膜自动分层中的应用

目的:对视网膜光学相干断层扫描图像中不同层和积液区域的分割。方法:提出一种基于深度学习的轻量级的神经网络,参考DRUNet体系、膨胀卷积和残差网络的架构,通过连接不同深度网络处得到的上采样输出,进行多尺度特征融合,使网络能够更好地识别出图像中的边界信息。结果:改进型DRUNet显著提升了视网膜分层的效果,准确率较U-Net提高了1.25%,同时能提前1~2次迭代达到传统U-Net的准确度。结论:本文采用的网络结构提高了对视网膜光学相干断层扫描图像的分割性能,同时降低了网络参数,具有强大的应用潜力。     

In vivo three-dimensional high-resolution imaging of rodent retina with spectral-domain optical coherence tomography

PURPOSE: To demonstrate the application of high-resolution spectral-domain optical coherence tomography (SD-OCT) for three-dimensional (3D) retinal imaging of small animals and quantitative retinal information extraction using 3D segmentation of the OCT images. METHODS: A high-resolution SD-OCT system was built for in vivo imaging of rodent retina. OCT fundus images similar to those acquired with a scanning laser ophthalmoscope (SLO) were constructed from the measured OCT data, which provided precise spatial registration of the OCT cross-sectional images on the fundus. A 3D segmentation algorithm was developed for calculation of the retinal thickness map. OCT images were compared by histologic examination. RESULTS: High-quality OCT images of the retinas of mice (B6/SJLF2 for normal retina, rhodopsin-deficient Rho(-/-) for photoreceptor degeneration, and LH(BETA)T(AG) for retinoblastoma) and rat (Wistar) were acquired. The OCT images compared well with histology. Not only was a 3D image of the tumor in a retinoblastoma mouse model successfully imaged in vivo but the tumor volume was extracted from the 3D image. Retinal thickness maps were calculated that enabled successful quantitative comparison of the retinal thickness distribution between the normal (202.3 +/- 9.3 microm) and the degenerative (102.7 +/- 12.6 microm) mouse retina. CONCLUSIONS: High-resolution spectral-domain OCT provides unprecedented high-quality 2D and 3D in vivo visualization of retinal structures of mouse and rat models of retinal diseases. With the capability of 3D quantitative information extraction and precise spatial registration, the OCT system made possible longitudinal study of ocular diseases that has been impossible to conduct.     

基于深度学习辅助诊断青光眼病灶检测算法及应用

目的研究基于卷积神经网络自动检测青光眼性视神经病变的深度学习算法,并探讨实现病灶区可视化的可行性。设计横断面研究。研究对象2014-2018年北京同仁医院5148例患者10296眼的眼底图像。方法在提供有无青光眼性视神经病变作为标记的前提下,首先基于ResNet深度模型训练一个深度神经网络,使用训练好的模型测试并计算其诊断分类的准确性。其次利用t-分布随机邻域嵌入可视化方法(t-SNE)实现对不同类别的深度特征分布可视化,生成相应的病灶区域热力图。计算该深度学习算法分类的敏感性、特异性和受试工作特性曲线下面积(AUC),并通过热力图评价其对某种类型病灶区的识别准确率以及对于诊断贡献最大的区域与专家的判别一致性。主要指标敏感性、特异性、AUC、识别准确率、判别一致性。结果在验证数据集中,该算法的AUC为0.996(95%CI,0.995-0.998),检测到病灶区的敏感性和特异性与受过培训的专业评分员相当(敏感性,96.2%vs 96.0%,P=0.76;特异性,97.7%vs 97.9%,P=0.81)。病灶区域热力图对视盘异常和盘沿丢失区域的识别准确率达到100%,对于诊断贡献最大的区域判别与青光眼专家的一致性达91.8%。结论运用深度学习算法检测青光眼性视神经病变的眼底图像具有较高的敏感性与特异性,同时基于t-SNE算法实现了对诊断贡献较大的病灶区域可视化。     

基于卷积神经网络的翼状胬肉病灶分割研究

目的：通过深度卷积神经网络方法对翼状胬肉病灶进行精准分割。

方法：在PSPNet模型结构的基础上构建Phase-fusion PSPNet网络结构用于翼状胬肉病灶的分割,该网络在金字塔池化模块后接入阶段上采样模块,以分阶段增大为原则逐步上采样,减少信息丢失,适合于边缘模糊的分割任务。将南京医科大学附属眼科医院提供的翼状胬肉眼表图像517张分为训练集(330张)、验证集(37张)、测试集(150张),其中训练集和验证集图像用于训练,测试集图像仅用于测试。比较翼状胬肉病灶智能分割和专家标注的结果。

结果：构建Phase-fusion PSPNet网络结构针对测试数据集的翼状胬肉病灶分割单类平均交并比(MIOU)和平均像素精确度(MPA)分别为86.31%和91.91%; 翼状胬肉单类交并比(IOU)和像素精确度(PA)分别为77.64%和86.10%。

结论：卷积神经网络可以实现翼状胬肉病灶的精准分割,有助于为医生进行进一步疾病诊断和手术建议提供重要参考,同时实现翼状胬肉智能诊断的可视化。     

OCT影像人工智能读片与医生读片对识别年龄相关性黄斑变性的一致性分析

目的：探讨在临床进行年龄相关性黄斑变性(ARMD)患者眼底光学相干断层扫描(OCT)图像人工智能(AI)读片的可行性。

方法：收集2019-11/2021-11 上海市静安区市北医院眼科门诊患者1 579眼OCT图像共1 579张,并收集眼科医生及AI的读片结果。通过Kappa值进行无ARMD和有ARMD分类结果的一致性分析。

结果：两名眼科医生之间在无ARMD和有ARMD读片结果的Kappa值为0.934; AI结果与眼科医生在无ARMD和有ARMD读片结果的Kappa值为0.738。AI对ARMD识别的灵敏度为73.08%,特异度为95.07%,曲线下区域面积(AUC)为0.841。

结论：AI在基于OCT图像的ARMD识别上与眼科医生有较高的一致性,适用于基层医院完成对ARMD的早期筛查和早期转诊工作。     

Quantification of photoreceptor layer thickness in normal eyes using optical coherence tomography

OBJECTIVE: To demonstrate the ability to segment and analyze individual intraretinal layers, including the outer retinal complex (ORC; outer nuclear layer and inner and outer segments of the photoreceptor cells), in healthy eyes using images acquired from the latest commercially available optical coherence tomography (OCT) system (StratusOCT; Carl Zeiss Meditec, Inc., Dublin, CA) and from the ultrahigh resolution OCT (UHR-OCT) prototype. METHODS: Thirty-seven eyes from 37 healthy subjects underwent complete ophthalmologic examination using StratusOCT and UHR-OCT. ORC was identified and measured using a segmentation algorithm. RESULTS: For StratusOCT, mean weighted ORC thickness +/- SD was 91.1 +/- 7.9 microm, and mean weighted total retinal thickness +/- SD was determined to be 258.9 +/- 10.1 microm. For UHR-OCT, mean weighted ORC thickness +/- SD was 96.4 +/- 6.3 microm, and mean weighted total retinal thickness +/- SD was determined to be 263.4 +/- 9.2 mum. There was a higher rate of algorithm failure with UHR-OCT images. CONCLUSIONS: Photoreceptor layer thickness can be calculated by measuring ORC on OCT images using a macular segmentation algorithm. ORC values may serve as a useful objective parameter in determining the efficacy of various therapeutic modalities that target the photoreceptor layer in various diseases.     

The interpretation of optical coherence tomography images of the retina.

PURPOSE: To determine the relationship between optical coherence tomography (OCT) images of the retina and retinal substructure in vitro and in vivo. METHODS: In vitro, OCT images of human and bovine retina were acquired after sequential excimer laser ablation of the inner retinal layers. Measurements of bands in the OCT images were compared with measurements of retinal layers on histology of the ablated specimens. In vivo, OCT images were acquired of retinal lesions in which there was a displacement of pigmented retinal pigment epithelial (RPE) cells: retinitis pigmentosa and laser photocoagulation (eight eyes each). RESULTS: The mean thickness of human inner OCT bands (131 microm; 95% confidence interval [CI], 122-140 microm) was 7.3 times that of the retinal nerve fiber layer (RNFL). This band persisted despite ablation greater than 140 microm. The inner aspect of the outer OCT band corresponded to the apical RPE, but the mean thickness of this band in human tissue (55 microm; 95% CI, 48-62 microm) was 2.6 times the thickness of the RPE-choriocapillaris complex. OCT measurement of total retinal thickness was accurate (coefficient of variance, 0.05) and precise (coefficient of correlation with light microscopy, 0.98). Hyperpigmented lesions gave rise to high signal, attenuating deeper signal; hypopigmented lesions had the opposite effect on deeper signal. CONCLUSIONS: The inner band is not RNFL specific, partly consisting of a surface-related signal. The location, not thickness, of the outer band corresponds to RPE melanin. Given the additional effect of polarization settings, precise OCT measurement of specific retinal layers is currently precluded.     

人工智能在眼部疾病中的应用及其挑战

随着人工智能技术的发展和普及,医学领域也出现了越来越多人工智能(AI)的身影。人工神经网络等新技术与临床的结合正成为研究热点,其中卷积神经网络(CNN)的深度学习算法在图像识别领域取得了巨大的成就,逐渐被用于糖尿病视网膜病变(DR)、年龄相关性黄斑变性(ARMD)、早产儿视网膜病变(ROP)、青光眼和白内障等多种眼科疾病的诊断和筛查中。目前针对不同眼科疾病,世界范围已有多个公开数据库,包括了眼底彩照、光学相干断层扫描(OCT)等多种图像资料,为眼科领域深度学习算法的训练和构建奠定了基础。同时算法本身也在不断优化,使相关AI产品的构建朝着更简便高效的方向发展,同时其临床运用也面临医学伦理和准入标准的问题。总之,深度学习算法的使用为几种常见眼科疾病的筛查诊断带来了巨大的改变也带来挑战,目前尚未大规模的投入临床应用中。本文针对人工智能在眼部疾病中的应用进展做综述,旨在总结这一领域的研究现状和现存问题,并提出对未来的展望。     

Spatial correlation of mouse photoreceptor-RPE thickness between SD-OCT and histology

Spectral Domain Optical Coherence Tomography (SD-OCT) applied to the mouse retina has been limited due to inherent movement artifacts and lack of resolution. Recently, SD-OCT scans from a commercially available imaging system have yielded retinal thickness values comparable to histology. However, these measurements are based on single point analysis of images. Here we report that using the Spectralis HRA + OCT Spectral Domain OCT and Fluorescein Angiography system (Heidelberg Engineering, Heidelberg, Germany), retinal thickness of linear expanses from SD-OCT data can be accurately assessed. This is possible by the development of a Spectralis-compatible ImageJ plug-in that imports 8-bit SLO and 32-bit OCT B-scan images, retaining scale and segmentation data and enabling analysis and 3D reconstruction. Moreover, mouse retinal layer thickness values obtained with this plug-in exhibit a high correlation to thickness measurements from histology of the same retinas. Thus, use of this ImageJ plug-in results in reliable quantification of long retinal expanses from in vivo SD-OCT images.     

深度学习在脉络膜分割中的应用研究进展

近年来,眼科作为高度依赖辅助成像的医学领域之一,一直处于深度学习算法应用的前沿。脉络膜的形态变化与眼底疾病的发生、发展以及治疗预后密切相关。光学相干断层扫描的快速发展极大地促进了对脉络膜形态和结构的精确分析。脉络膜分割及相关分析对于确定眼病的发病机制和治疗策略至关重要,然而,目前脉络膜主要依赖于繁琐、耗时和重复性低的人工手动分割。为了克服这些困难,近年来开发了用于脉络膜分割的深度学习方法,极大地提高了脉络膜分割的准确性和效率。本文旨在回顾不同眼病中脉络膜厚度的特征,探讨深度学习模型在测量脉络膜厚度中的最新应用及其优势,同时关注深度学习模型面临的挑战。