Refined Anam-Net:Lightweight Deep Learning Model for Improved Segmentation Performance of Optic Cup and Disc for Glaucoma Diagnosis

In this work,we aim to introduce some modifications to the Anam-Net deep neural network(DNN)model for segmenting optic cup(OC)and optic disc(OD)in retinal fundus images to estimate the cup-to-disc ratio(CDR).The CDR is a reliable measure for the early diagnosis of Glaucoma.In this study,we developed a lightweight DNN model for OC and OD segmentation in retinal fundus images.Our DNN model is based on modifications to Anam-Net,incorporating an anamorphic depth embedding block.To reduce computational complexity,we employ a fixed filter size for all convolution layers in the encoder and decoder stages as the network deepens.This modification significantly reduces the number of trainable parameters,making the model lightweight and suitable for resource-constrained applications.We evaluate the performance of the developed model using two publicly available retinal image databases,namely RIM-ONE and Drishti-GS.The results demonstrate promising OC segmentation performance across most standard evaluation metrics while achieving analogous results for OD segmentation.We used two retinal fundus image databases named RIM-ONE and Drishti-GS that contained 159 images and 101 retinal images,respectively.For OD segmentation using the RIM-ONE we obtain an f1-score(F1),Jaccard coefficient(JC),and overlapping error(OE)of 0.950,0.9219,and 0.0781,respectively.Similarly,for OC segmentation using the same databases,we achieve scores of 0.8481(F1),0.7428(JC),and 0.2572(OE).Based on these experimental results and the significantly lower number of trainable parameters,we conclude that the developed model is highly suitable for the early diagnosis of glaucoma by accurately estimating the CDR.

Association between polymorphism of optic disc related genes and susceptibility to primary open angle glaucoma in Inner Mongolia

Objective:To explore the relationship between the polymorphism of optic disc related genes and the susceptibility to primary open angle glaucoma(POAG)in Inner Mongolia.Methods:A retrospective study was adopted to include 108 patients who were diagnosed as POAG in six hospitals in Hohhot and Baotou from January of 2014 to December of 2016(POAG group).At the same time,120 healthy examinees were included in the control group.1-2 ml of whole blood was collected by use of EDTA anticoagulation tubes from each patient in these two groups.It was required to fully mix EDTA with whole blood in order to extract genomic DNA and place it in the-20℃ refrigerator.Mass spectrometry was used to identify the genotype of single nucleotide polymorphism(SNP)of RFTN1(rs690037),ATOH7(rs7916697,rs3858145),CDC7(rs1192415),CDKN2B(rs1063192)and SIX(rs10483727)in 108 POAG patients and 120 normal controls.χ^(2) test and binary logistic regression were used to analyze the relationship between the genetic polymorphism and the occurrence of POAG.Results:The frequency of G allele at CDKN2B(rs1063192)in POAG group was significantly higher than that in the control group(27%vs.17%),and the difference was of statistical significance[odds ratio(OR)=1.824,95%confidence interval(CI):1.163-2.861,p=.008];As to the comparison in the frequency of allele at the other 7 SNPs between the two groups,the differences were statistically significant(all p>.05).Additive and dominant models at rs1063192 indicated that the individuals with G allele were more susceptible to POAG,and the difference was of statistical significance(p<.05),and recessive models showed that the risk in the individuals with A allele was not significantly reduced,and the difference was of no statistical significance(p>.05).There was no statistically significant difference in the distribution of genotypes at other SNPs between POAG group and the control group(p>.05).Conclusions:The genetic polymorphism of CDKN2B(rs1063192)is associated with the susceptibility to POAG,and G allele may increase the risk for POAG.

Automated Detection of Optic Disc from Digital Retinal Fundus Images for Screening Systems of Diabetic Retinopathy

Optic disc(OD)detection is a main step while developing automated screening systems for diabetic retinopathy.We present a method to automatically locate and extract the OD in digital retinal fundus images.Based on the property that main blood vessels gather in OD,the method starts with Otsu thresholding segmentation to obtain candidate regions of OD.Consequently,the main blood vessels which are segmented in H channel of color fundus images in Hue saturation value(HSV)space.Finally,a weighted vessels’direction matched filter is proposed to roughly match the direction of the main blood vessels to get the OD center which is used to pick the true OD out from the candidate regions of OD.The proposed method was evaluated on a dataset containing 100 fundus images of both normal and diseased retinas and the accuracy reaches 98%.Furthermore,the average time cost in processing an image is 1.3 s.Results suggest that the approach is reliable,and can efficiently detect OD from fundus images.

Automatic Optic Disc Detection in Retinal Images Using FKMT‒MOPDF

In recent days,detecting Optic Disc(OD)in retinal images has been challenging and very important to the early diagnosis of eye diseases.The process of detecting the OD is challenging due to the diversity of color,intensity,brightness and shape of the OD.Moreover,the color similarities of the neighboring organs of the OD create difficulties during OD detection.In the proposed Fuzzy K‒Means Threshold(FKMT)and Morphological Operation with Pixel Density Feature(MOPDF),the input retinal images are coarsely segmented by fuzzy K‒means clustering and thresholding,in which the OD is classified from its neighboring organs with intensity similarities.Then,the segmented images are given as the input to morphological operation with pixel density feature calculations,which reduce the false detection in the small pixel of the OD.Finally,the OD area is detected by applying the Sobel edge detection method,which accurately detects the OD from the retinal images.After detection optimization,the proposed method achieved Sensitivity(SEN),Specificity(SPEC)and Accuracy(ACC),with 96.74%,96.78%and 96.92%in DiaretDB0(Standard Diabetic Retinopathy Database Calibration level 0),97.12%,97.10%and 97.75%in DiaretDB1(Standard Diabetic Retinopathy Database Calibration level 1)and 97.19%,97.47%and 97.43%in STARE(Structured Analysis of the Retina)dataset respectively.The experimental results demonstrated the proposed method’s efficiency for segmenting and detecting OD areas.

Automatic segmentation of optic disc and cup for CDR calculation

Glaucoma as an irreversible blinding opioid neuropathy disease, its blindness rate is the second only after cataract in the world. The optic cup-to-disc ratio(CDR) is generally considered to be an important clinical indicator for judging the severity of glaucoma by ophthalmologists from retinal fundus image. In this letter, we propose an automatic CDR measurement method that consists of a novel optic disc localization method and a simultaneous optic disc and cup segmentation network based on the improved U shape deep convolutional neural network. Experimental results demonstrate that the proposed method can achieve superior performance when compared with other existing methods. Thus, our method can be used as a powerful tool for glaucoma-assisted diagnosis.

Detection and Grading of Diabetic Retinopathy in Retinal Images Using Deep Intelligent Systems: A Comprehensive Review

Diabetic Retinopathy(DR)is an eye disease that mainly affects people with diabetes.People affected by DR start losing their vision from an early stage even though the symptoms are identified only at the later stage.Once the vision is lost,it cannot be regained but can be prevented from causing any further damage.Early diagnosis of DR is required for preventing vision loss,for which a trained ophthalmologist is required.The clinical practice is time-consuming and is not much successful in identifying DR at early stages.Hence,Computer-Aided Diagnosis(CAD)system is a suitable alternative for screening and grading of DR for a larger population.This paper addresses the different stages in CAD system and the challenges in identifying and grading of DR by analyzing various recently evolved techniques.The performance metrics used to evaluate the Computer-Aided Diagnosis system for clinical practice is also discussed.

基于边缘U^(2)-Net的视盘分割方法

彩色眼底图像中的视盘分割在识别眼科疾病中起着关键作用。针对因各种因素影响的视盘边缘分割不准确及分割算法效率低问题,提出一种基于轻量级U^(2)-Net、融入边缘注意力机制的视盘自动分割方法。该方法以轻量级U^(2)-Net为主干网络,使用视盘感兴趣区域提取的预处理方式去除无关特征,同时引入边缘注意力机制增强对视盘边缘特征的提取能力。在Drishti_GS和REFUGE两个公开数据集上的F1分数分别达到97.82%和97.36%,Dice相似系数分别达到97.15%和96.64%,IOU分别达到94.47%和93.50%,与其他网络模型相比表现出优越的分割性能,具有临床应用价值。

基于融合中间特征网络的视盘和视杯联合分割

针对视盘和视杯联合分割中视杯分割精度较差的问题,提出了一种融合编码与解码中间特征的U型网络(encode-decode middle feature fusion U-Net,EMFF-Net)。EMFF-Net使用预训练的ResNet34作为编码结构,在编码结构后加入密集空洞卷积和金字塔池化模块以产生复合感受域的特征,并使用交叉注意力连接替换U型网络结构中的跳跃连接。交叉注意力连接融合了编码特征与解码特征,通过通道注意力模块和空间注意力模块提取融合特征的信息用于强化解码特征,减小了解码特征与编码特征的语义沟壑。强化后的解码特征与编码特征再次融合后,通过解码结构输出视盘和视杯的联合分割结果。实验结果表明,与其他常用的分割方法相比,EMFF-Net的视盘和视杯联合分割效果较好,视杯分割性能有明显提升。

基于显著性引导的保凸性水平集分割算法

结合彩色眼底图像中视神经盘的解剖学特征,引入零水平集曲线的曲率变化来优化活动轮廓模型,提出了一种全自动分割视神经盘的水平集算法CRLSF。该算法利用显著性算法快速获取初始轮廓,且该初始轮廓与视神经盘轮廓非常接近,大幅度提高分割效率。采用保凸机制能够使得曲线在演化过程中不受到视神经盘周围毛细血管的影响以保持分割曲线的凸性。最后,在迭代中引入目标函数来评估算法的收敛性。实验表明,提出的CRLSF算法优于已有的算法。

基于ReSE-AG-Net的青光眼视盘视杯分割方法

分割视盘视杯是诊断青光眼的主要依据,因此提出了一种基于改进U-Net网络的青光眼视盘视杯分割方法,称为ReSE-AG-Net。并将ReSE-AG-Net、U-Net、M-Net、CE-Net和CDED-Net分别在Drishti-GS1数据集和RIM-ONE-R3数据集上进行测试,对得到分割后的视盘视杯的Dice系数和IoU进行对比。实验结果表明,在Drishti-GS1数据集上,分割视盘的Dice系数和IoU分别是0.989 1和0.959 0,分割视杯的Dice系数和IoU分别是0.931 0和0.871 0,在RIMONE-R3数据集上,分割视盘的Dice系数和IoU分别是0.978 9和0.958 7,分割视杯的Dice系数和IoU分别是0.863 7和0.760 1,对比U-Net模型和现有的分割模型均有一定的提升,证明了该方法的可行性。

基于对抗学习和引导机制的视盘和视杯联合分割

准确的视盘(OD)和视杯(OC)分割能够有效地辅助青光眼的诊断和监测,从而进一步提高治疗效果。然而,现有方法没有考虑到眼底图像不同通道之间的差异,并且难以实现对OC边界的精确分割。针对这个问题,提出一种基于对抗学习和引导机制的网络框架ALG-Net,旨在提高OD和OC的分割性能。ALG-Net由分割网络和鉴别器两部分组成。在分割网络中,构建引导融合模块(GFM),该模块将单通道特征信息与RGB图像特征融合,使网络充分学习眼底图像不同通道之间的差异信息,引导分割网络聚焦于关键区域。ALG-Net网络框架还采用了鉴别器,通过对抗学习的方式促进分割网络生成更真实的分割结果。在REFUGE和Drishti-GS数据集上进行广泛的实验评估,实验结果表明,ALG-Net在RUFUGE数据集上OD和OC分割的平衡精度分别达到了98.6%和95.9%,在Drishti-GS数据集上也表现出优异的性能。此外,ALG-Net的分割结果应用于青光眼分类任务,在RUFUGE数据集上ROC曲线下面积(AUC)为0.983,相较于经典UNet算法提高了0.015,为青光眼的早期诊断和监测提供了有力的支持。

基于改进U⁃Net的视盘视杯联合分割方法

青光眼是一种不可逆的致盲性眼疾,疾病早期症状不明显使得许多患者错失治疗的最佳时机。眼底照相作为最常见的青光眼筛查手段,眼底杯盘比值是诊断青光眼的重要指标之一。针对图像中视盘视杯分割精度不高的问题,构建了一种改进U⁃Net的视盘视杯联合分割模型CASSP⁃Net,引入CBAM注意力机制和空洞空间金字塔结构,进一步提升视盘视杯联合分割的精确度,在Drishti⁃GS和REFUGE数据集中进行测试,在Dice和IoU上分别获得92.03%和85.23%的较好表现。

正常人视盘、视杯形态及盘周神经纤维层光学相干断层定量检测

目的:观察光学相干断层扫描仪检测正常人视盘、视杯形态及盘周视网膜神经纤维层厚度。方法:应用CarlZeiss公司Humphrey2000型光学相干断层扫描仪检查系统对80例单侧眼病患者的80只正常对侧眼进行视盘"十字交叉"和"环形"扫描,分别测量视盘、视杯的直径,视杯深度,以及盘周视网膜神经纤维层厚度,计算杯盘比值及面积,统计各参数的平均值。结果:用视盘横纵扫描方法测量结果为:视盘纵扫直径为1.60±0.22mm,横扫直径为1.40±0.15mm,平均直径为1.50±0.22mm;视盘扫描平均面积为1.77±0.53mm2。视杯纵扫直径为0.54±0.22mm,视杯横扫直径为0.52±0.24mm,平均直径为0.54±0.23mm;视杯扫描平均面积为0.26±0.20mm2;正常人杯盘直径比纵扫为0.34±0.13,杯盘直径比横扫为0.38±0.15,平均为0.36±0.14;杯盘面积比平均为0.15±0.11。视杯深度纵扫为0.31±0.14mm,横扫为0.27±0.14mm,平均为0.29±0.14mm。用视盘周围环扫方法测量的盘周神经纤维层厚度结果:颞侧为95.57±17.63μm,上方为144.17±18.98μm,鼻侧为99.00±28.13μm,下方为155.53±18.65μm。结论:光学相干断层检测能做出视盘及视杯的定量测量和分析,得到的正常值范围可作为前部缺血性视神经病变、青光眼性视神经损害等视神经疾病临床观察的参考。

改进区域生长算法在视杯图像分割中的应用

目的:视杯图像分割对于通过眼底图像检测青光眼具有重要意义,在传统的区域生长算法基础上进行改进,提出了基于眼底图像的视杯自动检测分割方法.方法:首先,对眼底主要生理结构进行特征分析,为分割目标选取了绿色通道并根据阈值法粗略提取出感兴趣区域(ROI);其次,考虑到传统的区域生长算法在选取种子点时不精确、自适应性差等缺点,通过计算ROI的几何中心并结合中心亮度作为选取种子点的标准进行改进;最后,用5*5模板对眼底图像进行均值滤波,应用山谷差值准则和8邻域连通准则对眼底图像进行种子合并,最终准确分割出视杯.结果:应用这种方法,对高分辨率眼底图像(HRF)数据库中15张青光眼眼底图像和15张健康眼眼底图像逐张进行检测,准确率达到93.3%.结论:实验结果表明,该算法能快速、有效地自动检测出眼底图像中的视杯并将其正确的分割出来,与传统算法相比较该算法稳定可靠,有较高的分割灵敏度、特异度以及准确性.

眼底图像中视盘的平滑滤波与CV模型分割

可靠、有效的视盘分割对自动眼底病变诊断分析具有重要意义。提出了一种基于平滑滤波和CV模型的视盘分割算法。在利用投影法定位出视盘中心后,首先采用Gabor滤波技术提取并移除视盘局部区域的主血管,然后利用邻域信息通过插值运算填充被移除血管区域的像素,采用边缘保留平滑滤波对视盘区域进行平滑,最后利用基于区域信息的CV水平集模型分割出视盘的边界。该方法在MESSIDOR数据库上进行了测试,平均分割准确率为0.83,表现出了良好的分割性能。

人工智能眼底分析技术对青光眼病灶的诊断价值研究

背景青光眼是一种不可逆性的致盲性眼病,研究青光眼眼底图像特征,利用人工智能技术在专业临床数据库的基础上建立诊断模型,能够快速、客观地对青光眼患者眼底图像进行判断。目的评价一种基于人工智能眼底分析技术的青光眼病灶诊断系统的性能,并利用该系统探索青光眼疾病进展中的杯盘比发展情况。方法选取2020年3月-2021年4月4000例在解放军总医院第三医学中心完成眼底照相患者的眼底照片,将采集的眼底照片进行视杯、视盘及视网膜纤维层缺损的分割标注、病灶多标签的分类标注及青光眼分期标注,获得眼底照片-青光眼数据集。建立一个基于深度学习算法模型的人工智能青光眼病灶诊断系统,并随机选取眼底照片进行内部验证以评估系统性能。利用该人工智能系统对不同分期的青光眼眼底照片进行杯盘比识别计算,分析随青光眼疾病进展的杯盘比分布差异。结果本研究共应用了6837张眼底照片,其中60%(4102张)用作训练集,40%(2735张)用作验证集。在验证集中人工智能青光眼病灶诊断系统在视杯分割预测结果上召回率平均为0.837,精确度平均为0.814,交并比平均为0.816,AUC平均为0.874;在视盘分割预测结果上召回率平均为0.928,精确度平均为0.926,交并比平均为0.916,AUC平均为0.941;在视网膜神经纤维层缺损(retinal nerve fiber layer defect,RNFLD)分割预测结果上召回率平均为0.653,精确度平均为0.612,交并比平均为0.480,AUC平均为0.749。在验证集中人工智能青光眼病灶诊断系统在局限性RNFLD病灶预测结果上准确度平均为0.890,敏感度平均为0.896,特异性平均为0.638,AUC平均为0.893;在弥漫性RNFLD病灶预测结果上准确度平均为0.950,敏感度平均为0.744,特异性平均为0.961,AUC平均为0.901;在视盘出血病灶预测结果上准确度平均为0.966,敏感度平均为0.650,特异性平均为0.967,AUC平均为0.969;在视杯切迹病灶预测结果上准确度平均为0.951,敏感度平均为0.794,特异性平均为0.957,AUC平均为0.892。对不同分期的青光眼眼底照片进行杯盘比识别计算,杯盘比值随青光眼视神经病变的进展逐渐增大。结论将人工智能眼底分析技术应用于青光眼病灶诊断系统,可以为实现青光眼筛查提供思路。

基于神经网络的眼底视杯图象分割技术

根据眼底图象中视杯的边沿特征，提出了一种基于ＢＰ神经网络的视杯分割方法，详述了这种方法的实现过程，并研究了ＢＰ学习算法的加速问题．实验证明，这种方法的分割效果较好．

视盘相关基因多态性与内蒙古自治区原发性开角型青光眼易感性的关系

目的探讨视盘相关基因多态性与内蒙古自治区原发性开角型青光眼（POAG）易感性的关系。方法采用回顾性研究设计，纳入2014年1月至2016年12月内蒙古自治区（呼和浩特市、包头市）6所医院确诊的POAG患者108例作为POAG组，同期纳入120名健康体检者作为对照组。2个组患者用EDTA抗凝管抽取1—2ml全血，提取基因组DNA，充分混匀EDTA和全血，提取基因组DNA于-20℃冰箱中保存。采用质谱分析法对108例POAG患者和120例对照者的RFTN1（rs690037）、ATOH7（rs7916697、rs3858145）、CDC7（rs1192415）、CDKN2B（rs1063192）和SIX（rs10483727）单核苷酸多态性（SNP）位点进行基因分型，采用r检验和Logsitic二元回归分析各位点多态性与POAG发病的关系。结果POAG组CDKN2B（rs1063192）位点G等位基因频率高于对照组（27％与17％），差异有统计学意义[优势比（OR）=1．824，95％可信区间（CI）1．163～2．861，P=0．008]；2个组其他7个SNPs位点等位基因频率比较，差异均无统计学意义（均P〉0．05）。CDKN2B（rs1063192）位点的加性模型和显性模型显示，携带G等位基因个体罹患POAG风险增高，差异有统计学意义（P〈0．05），隐性模型显示，携带A等位基因个体罹患POAG风险无显著降低，差异无统计学意义（P〉0．05）。POAG和对照组其他SNPs各基因型分布比较，差异无统计学意义（P〉0．05）。结论CDKN2B（rs1063192）基因多态性与POAG易感性有关，G等位基因可增加个体罹患POAG的风险。

基于多特征融合的彩色眼底图像视杯分割方法

为了提高彩色眼底图像中视杯的分割精度,提出了一种基于多特征融合的彩色眼底图像视杯分割方法.首先提取感兴趣区域的血管;然后分割视盘区域,在视盘分割的基础上根据视杯的亮度特征采用模糊C均值聚类(FCM)法提取视杯候选区域,并根据视杯的形状和位置特征对候选区域依次进行镜像映射、椭圆拟合及椭圆校正,得到视杯的粗分割结果;最后利用杯沿的血管特征定位血管弯曲点,修正视杯粗分割结果,完成视杯的准确分割.对Glaucoma Repo眼底图像数据库进行测试,实验结果表明:该方法的灵敏度为87.15%,特异性为99.03%,准确率为98.12%,阳性预测值为82.03%,综合评价指标为84.51%,像素距离为18.80,具有较高的鲁棒性和有效性.

基于深度学习的视网膜病变眼底图视盘自动定位与分割研究

目的:研究视网膜病变眼底图像中视盘的自动定位与分割。方法:收集Messidor数据库中1200张眼底图,其中900张用于训练,300张用于测试;采用2个深度学习网络分别实现视盘在图像上的精确定位与分割,采用准确率、Dice系数、Jaccard系数、敏感性、特异性指标评估结果。结果:视盘定位准确率为100%,视盘分割准确率为97.38%,平均Dice系数、Jaccard系数、敏感性、特异性为0.9194、0.8549、0.9959、0.9700,平均处理时间为0.25 s。结论:深度学习方法能够实现高精度的视盘定位及视盘分割,有助于眼底图像的实时自动分析。