一种基于改进Attention U-net的联合视杯视盘分割方法

青光眼是当前世界范围内致盲的主要病因之一,其发病过程没有明显的特征。视杯盘比是青光眼诊断中最主要的评估指标之一,这使得视杯视盘的分割成为了目前青光眼诊断的关键。已有的视杯视盘分割方法大多基于手工提取的特征,低效且精度不高。提出一种名为MAR2U-net的深度神经网络架构用于青光眼视杯视盘的联合分割。它是基于Attention U-net的一种改进架构,通过在Attention U-net的基础之上引入递归残差卷积模块来提取更加深层次的特征,并结合多尺度的输入和多标签的Focal Tversky损失函数来提升模型的联合分割性能。实验结果表明,该方法在REFUGE数据集上的分割效果较已有方法取得了显著提升,为实现大规模的青光眼诊断筛查提供了基础。     

Water body segmentation of Synthetic Aperture Radar image using Deep Convolutional Neural Networks

Deep Convolutional Neural Networks are finding their way into modern machine learning tasks and proved themselves to become one of the best contenders for future development in the field. Several proposed methods in image segmentation and classification problems are giving us satisfactory results and could even perform better than humans in image recognition tasks. But also at the cost of their performance, they also require a huge amount of images for training and huge amount of computing power and time that makes them unrealistic in some situations where obtaining a large dataset is not feasible. In this work, an attempt is made for segmentation of Synthetic Aperture Radar (SAR) images which are not usually abundant enough for training, and are heavily affected by a kind of multiplicative noise called speckle noise. For the segmentation task, pre-defined filters are first applied to the images and are fed to hybrid CNN that is resulted from the concept of Inception and U-Net. The outcome of our proposed method has been examined for their effectiveness of application in a complete set of SAR images that are not used for training. The accuracy has also been compared with the manually annotated SAR images.     

结合通道注意力的特征融合多人姿态估计算法

为了提高二维复杂场景下多人姿态估计准确度和速度,提出了一种Mobile-YOLOv3模型与多尺度特征融合全卷积网络相结合的自顶向下多人姿态估计方法.利用深度可分离卷积改进YOLOv3网络以作为高效的人体目标检测器.针对网络特征下采样过程中上层高分辨率信息不断遗失问题,在经典U型网络结构中嵌入多尺度特征融合模块,从而使网络中的低尺度特征也包含高分辨率信息,并在特征融合模块中引入通道注意力机制,进一步突出多尺度融合特征图的关键通道信息.试验结果表明:相比于堆叠沙漏网络(Stacked Hourglass Network,SHN)和级联金字塔网络(Cascaded Pyramid Network,CPN),文中所提出的人体姿态估计算法在COCO数据集上的姿态估计平均准确率分别提高了4.7和3.7.     

风机叶片无人机红外热图像拼接方法

针对风机叶片红外图像拼接困难的问题,提出了一种基于无人机速度信息的风机叶片红外图像拼接方法。 首先,利用 U-net 网络预测获得叶片掩膜图像,从而去除冗余的背景信息;其次,计算平移、旋转、缩放参数使拼接图像配准;最后,使用 Multiband Blend 算法对拼接图像进行融合,消除视场与光照变化引起的拼缝。 实验结果表明,本文提出的方法在拼接处 x 梯度 方向上的 RMSE 小于 SURF 等传统图像拼接方法,拼接成功率达 97. 8%,并成功获取风机叶片红外全景图。 将 Multiband Blend 算法应用于叶片红外图像融合,结果表明融合后图像拼接处 RMSE 显著降低,过渡更加平滑。     

结合多分支结构和U-net的低照度图像增强

随着夜景拍摄技术的提高,低照度图像增强成为计算机视觉领域一个新的热点。但是由于光照不足、逆光、聚焦失败等因素的影响会导致光照强度不足,导致图像亮度和对比度过低。为了更好地处理低光照图像,提出了一种基于多分支结构和U-net结合的低照度图像增强算法。利用深度残差网络将图片不同层次的特征提取出来进行交叉合并。将得到的图像通过不同深度和结构的U-net进行增强。将U-net增强后的图像进行融合,最终得到了增强后的低照度图像。通过大量的实验表明,运用深度残差网络和U-net,可以更好地进行特征提取,低照度图像增强的效果也更好,很大程度上优于现有的技术。提出的方法不仅在视觉上提高了亮度和对比度,色彩更真实,更加符合人眼视觉系统特性,而且PSNR、SSIM等七项客观图像质量指标在几种算法中都是最优的。     

Ultrasonic evaluation of fetal lung development using deep learning with graph

BackgroundThe neonatal respiratory morbidity that was primarily caused by the immaturity of the fetal lung is an important clinical issue in close relation to the morbidity and mortality of the fetus. In clinics, the amniocentesis has been used to evaluate the fetal lung maturity, which is time-consuming, costly and invasive. As a non-invasive means, ultrasonography has been explored to quantitatively examine the fetal lung in the past decades. However, existing studies required the contour of the fetal lung which was delineated manually. This may lead to significant inter- and intra-observer variations.MethodsWe proposed a deep learning model for automated fetal lung segmentation and measurement, which was constructed combined U-Net with Graph model and pre-trained Vgg-16 network. The graph connection would extract stable feature for final segmentation and pre-trained method could speed up convergence.The model was trained with 3500 datasets augmented from 250 ultrasound images with both the fetal lung and heart delineated manually, and tested on 50 ultrasound images. In addition, the correlation between the size of fetal lung/heart as delineated by the model with gestational age was analyzed.ResultsThe fetal lung and cardiac area were segmented automatically with the accuracy, average Intersection over Union(IoU), sensitivity and precision being 0.991, 0.818, 0.909 and 0.888, respectively. In addition, the size of fetal lung/heart was well correlated with the gestational age, demonstrating good potentials for assessing the fetal development.ConclusionsThis study proposed a new robust method for automatic fetal lung segmentation in ultrasound images using Vgg16-GCN-UNet. Our proposed method could be utilized potentially not only to improve existing research in quantitative analyzing the fetal lung using ultrasound imaging technology, but also to alleviate the labor of the clinicians in routine measurement of the fetal lung/cardiac.     

基于轻量型U-net的钢材金相图像晶界分割方法

在金相组织的晶粒度自动化评估工作中,对晶粒边界识别的精准与否直接影响着金相组织晶粒度等级的评估准确度。针对钢材金相图像中晶粒边界密集程度高、边缘复杂且晶粒边界识别准确性低的问题,提出一种基于轻量型U-net卷积神经网络的金相图像晶界分割方法,该轻量型网络模型将浅层特征层用跳跃连接的方式拼接在上采样过程中,使网络学习到更多的有效特征信息;减少了网络层数并在特征提取过程中添加了一次卷积过程,减少了网络参数量并提高了对晶界的预测速度和准确率;实验结果表明,该方法在117张金相图像测试集上像素准确率达到93.91%、特异度为96.73%、灵敏度为81.6%。与传统U-net网络相比,像素准确率提高了0.2%,网络参数量相对减少了61.5%。本方法对金相晶界分割具有有效性和优越性。     

U-net与Dense-net相结合的视网膜血管提取

目的 视网膜血管健康状况的自动分析对糖尿病、心脑血管疾病以及多种眼科疾病的快速无创诊断具有重要参考价值。视网膜图像中血管网络结构复杂且图像背景亮度不均使得血管区域的准确自动提取具有较大难度。本文通过使用具有对称全卷积结构的U-net深度神经网络实现视网膜血管的高精度分割。方法 基于U-net网络中的层次化对称结构和Dense-net网络中的稠密连接方式,提出一种改进的适用于视网膜血管精准提取的深度神经网络模型。首先使用白化预处理技术弱化原始彩色眼底图像中的亮度不均,增强图像中血管区域的对比度;接着对数据集进行随机旋转、Gamma变换操作实现数据增广;然后将每一幅图像随机分割成若干较小的图块,用于减小模型参数规模,降低训练难度。结果 使用多种性能指标对训练后的模型进行综合评定,模型在DRIVE数据集上的灵敏度、特异性、准确率和AUC（area under the curve）分别达到0.740 9、0.992 9、0.970 7和0.917 1。所提算法与目前主流方法进行了全面比较,结果显示本文算法各项性能指标均表现良好。结论 本文针对视网膜图像中血管区域高精度自动提取难度大的问题,提出了一种具有稠密连接方式的对称全卷积神经网络改进模型。结果表明该模型在视网膜血管分割中能够达到良好效果,具有较好的研究及应用价值。     

SABOS-Net: Self-supervised attention based network for automatic organ segmentation of head and neck CT images

The segmentation of Organs At Risk (OAR) in Computed Tomography (CT) images is an essential part of the planning phase of radiation treatment to avoid the adverse effects of cancer radiotherapy treatment. Accurate segmentation is a tedious task in the head and neck region due to a large number of small and sensitive organs and the low contrast of CT images. Deep learning-based automatic contouring algorithms can ease this task even when the organs have irregular shapes and size variations. This paper proposes a fully automatic deep learning-based self-supervised 3D Residual UNet architecture with CBAM(Convolution Block Attention Mechanism) for the organ segmentation in head and neck CT images. The Model Genesis structure and image context restoration techniques are used for self-supervision, which can help the network learn image features from unlabeled data, hence solving the annotated medical data scarcity problem in deep networks. A new loss function is applied for training by integrating Focal loss, Tversky loss, and Cross-entropy loss. The proposed model outperforms the state-of-the-art methods in terms of dice similarity coefficient in segmenting the organs. Our self-supervised model could achieve a 4% increase in the dice score of Chiasm, which is a small organ that is present only in a very few CT slices. The proposed model exhibited better accuracy for 5 out of 7 OARs than the recent state-of-the-art models. The proposed model could simultaneously segment all seven organs in an average time of 0.02 s. The source code of this work is made available at https://github.com/seeniafrancis/SABOSNet .     

Gated Linear Model induced U-net for surrogate modeling and uncertainty quantification

We propose a novel deep learning based surrogate model for solving high-dimensional uncertainty quantification and uncertainty propagation problems. The proposed deep learning architecture is developed by integrating the well-known U-net architecture with the Gaussian Gated Linear Network (GGLN) and referred to as the Gated Linear Network induced U-net or GLU-net. The proposed GLU-net treats the uncertainty propagation problem as an image to image regression and hence, is extremely data efficient. Additionally, it also provides estimates of the predictive uncertainty. The network architecture of GLU-net is less complex with 44% fewer parameters than the contemporary works. We illustrate the performance of the proposed GLU-net in solving the Darcy flow problem under uncertainty under the sparse data scenario. We consider the stochastic input dimensionality to be up to 4225. Benchmark results are generated using the vanilla Monte Carlo simulation. We observe the proposed GLU-net to be accurate and extremely efficient even when no information about the structure of the inputs is provided to the network. Case studies are performed by varying the training sample size and stochastic input dimensionality to illustrate the robustness of the proposed approach.