带尺寸约束的弱监督眼底图像视盘分割

目的 医学图像的像素级标注工作需要耗费大量的人力。针对这一问题，本文以医学图像中典型的眼底图像视盘分割为例，提出了一种带尺寸约束的弱监督眼底图像视盘分割算法。方法 对传统卷积神经网络框架进行改进，根据视盘的结构特点设计新的卷积融合层，能够更好地提升分割性能。为了进一步提高视盘分割精度，本文对卷积神经网络的输出进行了尺寸约束，同时用一种新的损失函数对尺寸约束进行优化，所提的损失公式可以用标准随机梯度下降方法来优化。结果 在RIM-ONE视盘数据集上展开实验，并与经典的全监督视盘分割方法进行比较。实验结果表明，本文算法在只使用图像级标签的情况下，平均准确识别率（mAcc）、平均精度（mPre）和平均交并比（mIoU）分别能达到0.852、0.831、0.827。结论 本文算法不需要专家进行像素级标注就能够实现视盘的准确分割，只使用图像级标注就能够得到像素级标注的分割精度。缓解了医学图像中像素级标注难度大的问题。     

Ant Colony Optimization-based method for optic cup segmentation in retinal images

An accurate detection of the cup region in retinal images is necessary to obtain relevant measurements for glaucoma detection. In this work, we present an Ant Colony Optimization-based method for optic cup segmentation in retinal fundus images. The artificial agents will construct their solutions influenced by a heuristic that combines the intensity gradient of the optic disc area and the curvature of the vessels. On their own, the exploration capabilities of the agents are limited; however, by sharing the experience of the entire colony, they are capable of obtaining accurate cup segmentations, even in images with a weak or non-obvious pallor. This method has been tested with the RIM-ONE dataset, yielding an average overlapping error of 24.3% of the cup segmentation and an area under the curve (AUC) of 0.7957 using the cup to disc ratio for glaucoma assessment.     

视觉显著性的眼底图像视盘检测

目的 青光眼是导致失明的主要疾病之一,视盘区域的形状、大小等参数是青光眼临床诊断的重要指标。然而眼底图像通常亮度低、对比度弱,且眼底结构复杂,各组织以及病灶干扰严重。为解决上述问题,实现视盘的精确检测,提出一种视觉显著性的眼底图像视盘检测方法。方法 首先,依据视盘区域显著的特点,采用一种基于视觉显著性的方法对视盘区域进行定位;其次,采用全卷积神经网络（fully convolutional neural network,FCN）预训练模型提取深度特征,同时计算视盘区域的平均灰度,进而提取颜色特征;最后,将深度特征、视盘区域的颜色特征和背景先验信息融合到单层元胞自动机（single-layer cellular automata,SCA）中迭代演化,实现眼底图像视盘区域的精确检测。结果 在视网膜图像公开数据集DRISHTI-GS、MESSIDOR和DRIONS-DB上对本文算法进行实验验证,平均相似度系数分别为0.965 8、0.961 6和0.971 1;杰卡德系数分别为0.934 1、0.922 4和0.937 6;召回率系数分别为0.964 8、0.958 9和0.967 4;准确度系数分别为0.996 6、0.995 3和0.996 8,在3个数据集上均可精确地检测视盘区域。实验结果表明,本文算法精确度高,鲁棒性强,运算速度快。结论 本文算法能够有效克服眼底图像亮度低、对比度弱及血管、病灶等组织干扰的影响,在多个视网膜图像公开数据集上进行验证均取得了较好的检测结果,具有较强的泛化性,可以实现视盘区域的精确检测。     

无监督混阶栈式稀疏自编码器的图像分类学习

目前多数图像分类的方法是采用监督学习或者半监督学习对图像进行降维,然而监督学习与半监督学习需要图像携带标签信息。针对无标签图像的降维及分类问题,提出采用混阶栈式稀疏自编码器对图像进行无监督降维来实现图像的分类学习。首先,构建一个具有三个隐藏层的串行栈式自编码器网络,对栈式自编码器的每一个隐藏层单独训练,将前一个隐藏层的输出作为后一个隐藏层的输入,对图像数据进行特征提取并实现对数据的降维。其次,将训练好的栈式自编码器的第一个隐藏层和第二个隐藏层的特征进行拼接融合,形成一个包含混阶特征的矩阵。最后,使用支持向量机对降维后的图像特征进行分类,并进行精度评价。在公开的四个图像数据集上将所提方法与七个对比算法进行对比实验,实验结果表明,所提方法能够对无标签图像进行特征提取,实现图像分类学习,减少分类时间,提高图像的分类精度。     

融合上下文和注意力的视盘视杯分割

目的 青光眼会对人的视力造成不可逆的损伤,从眼底图像中精确地分割视盘和视杯是青光眼诊治中的一项重要工作,为有效提升视盘和视杯的分割精度,本文提出了融合上下文和注意力的视盘视杯分割方法（context attention U-Net,CA-Net）。方法 进行极坐标转换,在极坐标系下进行分割可以平衡数据分布。使用修改的预训练ResNet作为特征提取网络,增强特征提取能力。采用上下文聚合模块（context aggregation module,CAM）多层次聚合图像上下文信息,使用注意力指导模块（attention guidance module,AGM）对融合后的特征图进行特征重标定,增强有用特征;使用深度监督思想同时对浅层网络权重进行训练,同时在视杯分割网络中引入了先验知识,约束对视杯的分割。结果 在3个数据集上与其他方法进行对比实验,在Drishti-GS1数据集中,分割视盘的Dice （dice coefficient）和IOU （intersection-over-union）分别为0.981 4和0.963 5,分割视杯的Dice和IOU分别为0.926 6和0.863 3;在RIM-ONE （retinal image database for optic nerve evaluation）-v3数据集中,分割视盘的Dice和IOU分别为0.976 8和0.954 6,分割视杯的Dice和IOU分别为0.864 2和0.760 9;在Refuge数据集中,分割视盘的Dice和IOU分别为0.975 8和0.952 7,分割视杯的Dice和IOU分别为0.887 1和0.797 2,均优于对比算法。同时,消融实验验证了各模块的有效性,跨数据集实验进一步表明了CA-Net的泛化性,可视化图像也表明CA-Net能够分割出更接近标注的分割结果。结论 在Drishti-GS1、RIM-ONE-v3和Refuge三个数据集的测试结果表明,CA-Net均能取得最优的视盘和视杯分割结果,跨数据集测试结果也更加表明了CA-Net具有良好的泛化性能。     

A real time cloud-based framework for glaucoma screening using EfficientNet

These days one of the major causes of partial or complete blindness that has affected a majority of people all around the world is glaucoma. Glaucoma is caused as a result of increased fluid pressure inside the optic nerves called intra ocular pressure. A real time cloud-based framework for screening the glaucoma suspect’s retinal fundus images as received by the people on the public cloud, is proposed in this paper. In the proposed framework the existence of glaucoma and analysis of the retinal fundus images is achieved by deep learning technique and convolutional neural network respectively. EfficientNet and UNet++ models are used to identify the presence of glaucoma. On comparing our framework to various state-of-the-art models and quantitative assessment are performing on various benchmark datasets like RIM-ONE and DRISHTI-GS1, it was found that the proposed framework is scalable, location independent, and easily accessible to one and all due to the cloud platform.

     

ASRNet: Adversarial Segmentation and Registration Networks for Multispectral Fundus Images

Multispectral imaging (MSI) technique is often used to capture images of the fundus by illuminating it with different wavelengths of light. However, these images are taken at different points in time such that eyeball movements can cause misalignment between consecutive images. The multispectral image sequence reveals important information in the form of retinal and choroidal blood vessel maps, which can help ophthalmologists to analyze the morphology of these blood vessels in detail. This in turn can lead to a high diagnostic accuracy of several diseases. In this paper, we propose a novel semi-supervised end-to-end deep learning framework called “Adversarial Segmentation and Registration Nets” (ASRNet) for the simultaneous estimation of the blood vessel segmentation and the registration of multispectral images via an adversarial learning process. ASRNet consists of two subnetworks: (i) A segmentation module S that fulfills the blood vessel segmentation task, and (ii) A registration module R that estimates the spatial correspondence of an image pair. Based on the segmention-driven registration network, we train the segmentation network using a semi-supervised adversarial learning strategy. Our experimental results show that the proposed ASRNet can achieve state-of-the-art accuracy in segmentation and registration tasks performed with real MSI datasets.     

基于医学影像分割方法的多模态语料库构建

电子文本病历语料库可提供相关医学影像的定性诊断结果,但缺乏直观影像和文本标注信息,不利于有效管理医学数据和医科学生自主学习相关医学知识.针对此问题,文中提出基于深度水平集算法的医学影像分割方法,对医学影像进行自动分割,给出感兴趣区域的轮廓结果及相关定量指标,并结合自然语言处理方法实现电子病历文本的标注,增强影像与文本病历多模态语料库的信息表征能力.在青光眼影像数据上的实验表明,文中方法可精准分割眼底图像中视盘和视杯,有效构建具有直观影像标记与对应病历文本的多模态语料库.     

基于生成对抗网络的自动细胞核分割半监督学习方法

为了减少对标注图像数量的依赖,提出一种新颖的半监督学习方法用于细胞核的自动分割。首先,通过新的卷积神经网络（CNN）从背景中自动提取细胞区域。其次,判别器网络通过应用全卷积网络来为输入的图像生成置信图;同时耦合对抗性损失和标准交叉熵损失,以改善分割网络的性能。最后,将标记图像和无标记图像与置信图结合来训练分割网络,使分割网络可以在提取的细胞区域中识别单个细胞核。对84张图像（训练集中的1/8图像带标注,其余图像无标注）的实验结果表明,提出的细胞核分割方法的分割准确率度量（SEG）得分可以达到77.9%,F1得分可以达到76.0%,这比该方法使用670张图像且训练集中的所有图像都带标注时的表现要好。     

基于生成对抗网络的自动细胞核分割半监督学习方法

为了减少对标注图像数量的依赖，提出一种新颖的半监督学习方法用于细胞核的自动分割。首先，通过新的卷积神经网络（CNN）从背景中自动提取细胞区域。其次，判别器网络通过应用全卷积网络来为输入的图像生成置信图；同时耦合对抗性损失和标准交叉熵损失，以改善分割网络的性能。最后，将标记图像和无标记图像与置信图结合来训练分割网络，使分割网络可以在提取的细胞区域中识别单个细胞核。对84张图像（训练集中的1/8图像带标注，其余图像无标注）的实验结果表明，提出的细胞核分割方法的分割准确率度量（SEG）得分可以达到77.9%，F1得分可以达到76.0%，这比该方法使用670张图像且训练集中的所有图像都带标注时的表现要好。     

基于单类分类器的半监督学习

提出一种结合单类学习器和集成学习优点的Ensemble one-class半监督学习算法.该算法首先为少量有标识数据中的两类数据分别建立两个单类分类器.然后用建立好的两个单类分类器共同对无标识样本进行识别,利用已识别的无标识样本对已建立的两个分类面进行调整、优化.最终被识别出来的无标识数据和有标识数据集合在一起训练一个基分类器,多个基分类器集成在一起对测试样本的测试结果进行投票.在5个UCI数据集上进行实验表明,该算法与tri-training算法相比平均识别精度提高4.5%,与仅采用纯有标识数据的单类分类器相比,平均识别精度提高8.9%.从实验结果可以看出,该算法在解决半监督问题上是有效的.     

不完备弱标记数据的粗糙协同学习模型

针对不完备弱标记数据的学习问题,提出基于粗糙集理论的半监督协同学习模型.首先定义不完备弱标记数据的半监督差别矩阵,提出充分、具有差异性的约简子空间获取算法.然后在有标记数据集上利用各约简子空间训练两个基分类器.在无标记数据上,各分类器基于协同学习的思想标注信度较大的无标记样本给另一分类器学习,迭代更新直至无可利用的无标记数据.UCI数据集实验对比分析表明,文中模型可以获得更好的不完备弱标记数据的分类学习性能,具有有效性.     

基于半监督学习的增量图像分类方法

为有效使用大量未标注的图像进行分类,提出一种基于半监督学习的图像分类方法。通过共同的隐含话题桥接少量已标注的图像和大量未标注的图像,利用已标注图像的Must-link约束和Cannot-link约束提高未标注图像分类的精度。实验结果表明,该方法有效提高Caltech-101数据集和7类图像集约10%的分类精度。此外,针对目前绝大部分半监督图像分类方法不具备增量学习能力这一缺点,提出该方法的增量学习模型。实验结果表明,增量学习模型相比无增量学习模型提高近90%的计算效率。关键词半监督学习,图像分类,增量学习中图法分类号TP391。41IncrementalImageClassificationMethodBasedonSemi-SupervisedLearningLIANGPeng^1,2,LIShao-Fa²,QINJiang-Wei²,LUOJian-Gao³¹(SchoolofComputerScienceandEngineering,GuangdongPolytechnicNormalUniversity,Guangzhou510665)²(SchoolofComputerScienceandEngineering,SouthChinaUniversityofTechnology,Guangzhou510006)³(DepartmentofComputer,GuangdongAIBPolytechnicCollege,Guangzhou510507)ABSTRACTInordertouselargenumbersofunlabeledimageseffectively,animageclassificationmethodisproposedbasedonsemi-supervisedlearning。Theproposedmethodbridgesalargeamountofunlabeledimagesandlimitednumbersoflabeledimagesbyexploitingthecommontopics。Theclassificationaccuracyisimprovedbyusingthemust-linkconstraintandcannot-linkconstraintoflabeledimages。TheexperimentalresultsonCaltech-101and7-classesimagedatasetdemonstratethattheclassificationaccuracyimprovesabout10%bytheproposedmethod。Furthermore,duetothepresentsemi-supervisedimageclassificationmethodslackingofincrementallearningability,anincrementalimplementationofourmethodisproposed。Comparingwithnon-incrementallearningmodelinliterature,theincrementallearningmethodimprovesthecomputationefficiencyofnearly90%。     

主动协同半监督粗糙集分类模型

粗糙集理论是一种有监督学习模型,一般需要适量有标记的数据来训练分类器。但现实一些问题往往存在大量无标记的数据,而有标记数据由于标记代价过大较为稀少。文中结合主动学习和协同训练理论,提出一种可有效利用无标记数据提升分类性能的半监督粗糙集模型。该模型利用半监督属性约简算法提取两个差异性较大的约简构造基分类器,然后基于主动学习思想在无标记数据中选择两分类器分歧较大的样本进行人工标注,并将更新后的分类器交互协同学习。UCI数据集实验对比分析表明,该模型能明显提高分类学习性能,甚至能达到数据集的最优值。     

Software quality estimation with limited fault data: a semi-supervised learning perspective

We addresses the important problem of software quality analysis when there is limited software fault or fault-proneness data. A software quality model is typically trained using software measurement and fault data obtained from a previous release or similar project. Such an approach assumes that fault data is available for all the training modules. Various issues in software development may limit the availability of fault-proneness data for all the training modules. Consequently, the available labeled training dataset is such that the trained software quality model may not provide predictions. More specifically, the small set of modules with known fault-proneness labels is not sufficient for capturing the software quality trends of the project. We investigate semi-supervised learning with the Expectation Maximization (EM) algorithm for software quality estimation with limited fault-proneness data. The hypothesis is that knowledge stored in software attributes of the unlabeled program modules will aid in improving software quality estimation. Software data collected from a large NASA software project is used during the semi-supervised learning process. The software quality model is evaluated with multiple test datasets collected from other NASA software projects. Compared to software quality models trained only with the available set of labeled program modules, the EM-based semi-supervised learning scheme improves generalization performance of the software quality models.     

基于自编码神经网络的半监督联邦学习模型

联邦学习是一种新型的分布式机器学习方法,可以使得各客户端在不分享隐私数据的前提下共同建立共享模型。然而现有的联邦学习框架仅适用于监督学习,即默认所有客户端数据均带有标签。由于现实中标记数据难以获取,联邦学习模型训练的前提假设通常很难成立。为解决此问题,对原有联邦学习进行扩展,提出了一种基于自编码神经网络的半监督联邦学习模型ANN-SSFL,该模型允许无标记的客户端参与联邦学习。无标记数据利用自编码神经网络学习得到可被分类的潜在特征,从而在联邦学习中提供无标记数据的特征信息来作出自身贡献。在MNIST数据集上进行实验,实验结果表明,提出的ANN-SSFL模型实际可行,在监督客户端数量不变的情况下,增加无监督客户端可以提高原有联邦学习精度。     

An automated CNN architecture search for glaucoma diagnosis based on NEAT

Glaucoma is an ocular disease that causes damage to the optic nerve, inducing successive narrowing of the visual field in affected patients due to an increased intraocular pressure, which can lead patients to blindness in an advanced stage without clinical reversal. For several years, the use of deep learning with convolutional neural networks (CNNs) has been successfully put into practice for several years. However, building a deep learning network requires an amount of experiments to find the fittest parameters, best choice of layers and an amount of available data. Thus it is not always able to produce satisfactory results due to the amount of parameters that need to be configured to adapt the CNN architecture to the problem in question, in most cases, with small datasets. Based on this scenario, this paper proposes and analyzes a CNN architecture construction from scratch, based on Neuroevolution of Augmenting Topologies for diagnosing glaucoma from fundus images. The method was evaluated with RIM-ONE and the combination of five glaucoma datasets in which we highlight 0.961 and 0.943 of f1-score, respectively for each dataset.

     

A new graph-based semi-supervised method for surface defect classification

Vision-based defect classification is an important technology to control the quality of product in manufacturing system. As it is very hard to obtain enough labeled samples for model training in the real-world production, the semi-supervised learning which learns from both labeled and unlabeled samples is more suitable for this task. However, the intra-class variations and the inter-class similarities of surface defect, named as the poor class separation, may cause the semi-supervised methods to perform poorly with small labeled samples. While graph-based methods, such as graph convolution network (GCN), can solve the problem well. Therefore, this paper proposes a new graph-based semi-supervised method, named as multiple micrographs graph convolutional network (MMGCN), for surface defect classification. Firstly, MMGCN performs graph convolution by constructing multiple micrographs instead of a large graph, and labels unlabeled samples by propagating label information from labeled samples to unlabeled samples in the micrographs to obtain multiple labels. Weighting the labels can obtain the final label, which can solve the limitations of computation complexity and practicality of original GCN. Secondly, MMGCN divides unlabeled dataset into multiple batches and sets an accuracy threshold. When the model accuracy reaches the threshold, the unlabeled datasets are labeled in batches. A famous case has been used to evaluate the performance of the proposed method. The experimental results demonstrate that the proposed MMGCN can achieve better computation complexity and practicality than GCN. And for accuracy, MMGCN can also obtain the best performance and the best class separation in the comparison with other semi-supervised surface defect classification methods.     

Automated segmentation of optic disc region on retinal fundus photographs: Comparison of contour modeling and pixel classification methods

The automatic determination of the optic disc area in retinal fundus images can be useful for calculation of the cup-to-disc (CD) ratio in the glaucoma screening. We compared three different methods that employed active contour model (ACM), fuzzy c-mean (FCM) clustering, and artificial neural network (ANN) for the segmentation of the optic disc regions. The results of these methods were evaluated using new databases that included the images captured by different camera systems. The average measures of overlap between the disc regions determined by an ophthalmologist and by using the ACM (0.88 and 0.87 for two test datasets) and ANN (0.88 and 0.89) methods were slightly higher than that by using FCM (0.86 and 0.86) method. These results on the unknown datasets were comparable with those of the resubstitution test; this indicates the generalizability of these methods. The differences in the vertical diameters, which are often used for CD ratio calculation, determined by the proposed methods and based on the ophthalmologist's outlines were even smaller than those in the case of the measure of overlap. The proposed methods can be useful for automatic determination of CD ratios.     

A semi-supervised learning detection method for vision-based monitoring of construction sites by integrating teacher-student networks and data augmentation

Recently, deep-learning detection methods have achieved huge success in the vision-based monitoring of construction sites in terms of safety control and productivity analysis. However, deep-learning detection methods require large-scale datasets for training purposes, and such datasets are difficult to develop due to the limited accessibility of construction images and the need for labor-intensive annotations. To address this problem, this research proposes a semi-supervised learning detection method for construction site monitoring based on teacher–student networks and data augmentation. The proposed method requires a limited number of labeled data to achieve high detection performance in construction scenarios. Initially, the proposed method trains the teacher object detector with labeled data following weak data augmentation. Next, the trained teacher object detector generates pseudo-detection results from unlabeled images that have been weakly augmented. Finally, the student object detector is trained with the pseudo-detection results and unlabeled images that have been both weakly and strongly augmented. In our experiments, 10,000 annotated construction images from the Alberta Construction Image Dataset (ACID) have been divided into a training set (70%) and a validation set (30%). The proposed method achieved a 91% mean average precision (mAP) on the validation set while only requiring 30% of the training set. In comparison, the existing supervised learning method ResNet50 Faster R-CNN achieved a mAP of 90.8% when training on the full training set. These experimental results show the potential of the proposed method in terms of reducing the time, effort, and costs spent on developing construction datasets. As such, this research has explored the potential of semi-supervised learning methods and increased the practicality of vision-based monitoring systems in the construction industry.