Automatic segmentation and plaque characterization in atherosclerotic carotid artery MR images

In vivo MRI provides a means to non-invasively image and assess the morphological features of atherosclerotic carotid arteries. To assess quantitatively the degree of vulnerability and the type of plaque, the contours of the lumen, outer boundary of the vessel wall and plaque components, need to be traced. Currently this is done manually, which is time-consuming and sensitive to inter- and intra-observer variability. The goal of this work was to develop an automated contour detection technique for tracing the lumen, outer boundary and plaque contours in carotid MR short-axis black-blood images. Seventeen patients with carotid atherosclerosis were imaged using high-resolution in vivo MRI, generating a total of 50 PD- and T1-weighted MR images. These images were automatically segmented using the algorithm presented in this work, which combines model-based segmentation and fuzzy clustering to detect the vessel wall, lumen and lipid core boundaries. The results demonstrate excellent correspondence between automatic and manual area measurements for lumen (r=0.92) and outer (r=0.91), and acceptable correspondence for fibrous cap thickness (r=0.71). Though further optimization is required, our algorithm is a powerful tool for automatic detection of lumen and outer boundaries, and characterization of plaque in atherosclerotic vessels.     

面向安检X光图像的违禁物品语义分割与识别研究

针对安检X光图像中违禁物品大小不一、物品摆放随意且存在重叠遮挡的技术难题,提出了一种改进的HRNet多尺度特征融合网络模型,实现图像中违禁物品的自动分割与识别。在编码阶段,利用HRNet网络中的多分辨率并行网络架构,提取多尺度特征,解决安检X光图像违禁物品尺度多样化的问题。在解码阶段,提出一种多层级特征聚合模块,采用数据相关上采样方法减少信息丢失,并聚合编码阶段提取的特征,以对物品进行更完整表征。在网络整体架构中,嵌入基于注意力机制的去遮挡模块加强模型的边缘感知能力,缓解安检X光图像中物品重叠遮挡严重的问题,提高模型的分割识别精度。通过在PIDray安检图像公开数据集进行实验,结果表明,在Easy、Hard、Hidden 3个验证子集上分别取得了73.15%、69.47%、58.33%的平均交并比,相比原始HRNet模型,分别提升了0.49%、1.17%、5.69%,总体平均交并比提升约2.45%。     

Automatic segmentation of coronary tree in CT angiography images

In this study, we proposed an automatic approach to segment the three‐dimensional coronary tree from computed tomography angiography (CTA) image data. Compared to the previous studies, our approach can efficiently locate the coronary root through extracting aorta by using circular Hough transform. Then, we extracted the two‐dimensional coronary borders in multiple projection directions by dynamic programming and used the border information to reconstruct the coronary surface by the spline fitting. The performance of our approach has been validated on the CTA dataset from 50 subjects, with the comparison to the manual segmentation by an experienced medical physician on the corresponding CTA data and X‐ray angiography data. Our experiments have shown that the average bias between our approach and the manual segmentation are 2.44 mm² with confidence interval [‐2.20 mm², 6.68 mm²] for cross‐sectional lumen area, 0.84 mm with confidence interval [‐2.76 mm, 4.44 mm] for maximum lumen diameter, and 0.32 mm with confidence interval [‐1.14 mm, 1.78 mm] for minimum lumen diameter. The results can demonstrate the potential effectiveness of our approach in CTA image processing.     

Automated segmentation and volumetric analysis of renal cortex,medulla, and pelvis based on non-contrast-enhanced T1- and T2-weighted MR images

Object

The aim of our study was to enable automatic volumetry of the entire kidneys as well as their internal structures (cortex, medulla, and pelvis) from native magnetic resonance imaging (MRI) data sets.

Materials and methods

Segmentation of the entire kidneys and differentiation of their internal structures were performed in 12 healthy volunteers based on non-contrast-enhanced T1- and T2-weighted MR images. Two data sets (each acquired in one breath-hold) were co-registered using a rigid registration algorithm compensating for possible breathing-related displacements. An automatic algorithm based on thresholding and shape detection segmented the kidneys into their compartments and was compared to a manual labeling procedure.

Results

The resulting kidney volumes of the automated segmentation correlated well with those created manually (R ² = 0.96). Average volume errors were determined to be 4.97 ± 4.08 % (entire kidney parenchyma), 7.03 ± 5.56 % (cortex), 12.33 ± 7.35 % (medulla), and 17.57 ± 14.47 % (pelvis). The variation of the kidney volume resulting from the automatic algorithm was found to be 4.76 % based on the measuring of one volunteer with three independent examinations.

Conclusion

The results demonstrate the feasibility of an accurate and repeatable automatic segmentation of the kidneys and their internal structures from non-contrast-enhanced magnetic resonance images.     

一种光学遥感图像海陆分割方法

海陆分割是近岸舰船检测中的重要步骤之一,为解决有波浪的光学遥感图像近岸海陆分割问题,确保其性能的准确度和鲁棒性,提出了一种新的近岸海陆分割方法。该方法首先使用均值漂移(Mean Shift)算法将图像分割成若干同质区域,然后与OTSU分割结果相结合,采用连通性区域分析的方法提取出海水区域,最后去除阴影区域,得到海陆分割的结果。实验证明,该方法能够处理海水中存在波浪的近岸舰船检测问题、抗干扰性好、准确度高、鲁棒性强。     

基于背景色彩和PCNN的磨粒图像单通道分割

为有效地从磨粒图像中分离出磨粒区域,根据彩色磨粒图像背景相对单一的特点,提出了一种基于背景色彩和脉冲耦合神经网络(PCNN)的磨粒图像单通道分割方法.首先,利用Otsu算法对磨粒图像进行预分割处理,并计算图像背景色彩的统计特征,以此作为背景的特征色彩矢量;然后,引入综合色距函数,通过计算图像中各像素与背景特征色彩之间的色差,将彩色磨粒图像三通道问题转化成单通道问题;最后,利用简化PCNN对构造的色差矩阵进行分割,从而得到彩色磨粒图像的分割结果.实验结果表明,该方法对含有单个磨粒和多个磨粒的彩色磨粒图像均能进行精确自动分割,是一种比较理想的分割方法.     

面向地理对象的高分光学与SAR影像一体化分割

由于光学和SAR影像成像方式的巨大差异,为提取统一的光学和SAR对象集合造成了很大的困难。为此,本文提出了一种面向地理对象的高分光学与SAR的一体化分割方法。该方法不同于传统一体化分割同时处理异源影像的策略,而仅对光学影像进行分割,从而获得可靠的地理对象集合;在此基础上,为每个对象自适应提取标记点,并依据粗配准结果将其投影到SAR影像中;最后,在SAR影像中开展基于标记点的区域增长,最终获得与光学影像分割对象相匹配的对象集合。通过对多组光学和SAR影像的实验结果表明,所提出方法提取的光学—SAR匹配对象集合更加接近实际的地理对象,且J-value可达7.8以上,在目视分析和定量评价中均显著优于对比方法。     

Analysis of contrast-enhanced MR images to assess renal function

The image analysis and kinetic modeling methods used in dynamic contrast-enhanced magnetic resonance imaging of the kidney are reviewed. Image analysis includes various techniques of coregistration and segmentation. Few methods have been completely implemented. Nevertheless, the use of coregistration may become a standard to decrease the effect of motion on abdominal images and improve the quality of the renal signals. Kinetic models are classified into three categories: enhancement-based, external and internal representations. Enhancement-based representations are limited to a basic analysis of the tracer concentration curves in the kidneys. Their relationship to the underlying physiology is complex and undefined. However, they can be used to evaluate the split renal function. External representations assess the kidney input and output. An external representation based on the up-slope of the renal enhancement to calculate the renal perfusion is commonly used because of its simplicity. In contrast, external representation based on deconvolution or identification methods remain underexploited. For glomerular filtration, an internal representation based on a two-compartmental model is mostly used. Internal representations based on multi-compartmental models describe the renal function in a more realistic way. Because of their numerical complexity, these models remain rarely used.     

相差细胞图像的光照均衡与分割计数方法研究

为了解决相差显微细胞图像的光照不均匀,实现准确、快速的细胞图像分割与计数,本文提出了一种图像光照均衡,并结合双阈值分割计数的方法。该方法对图像进行分块,依据子块与整幅图像平均灰度的比值进行亮度调节,并采用高斯函数权值优化比值,消除因图像分块而产生的“方块效应”,并通过图像熵与高斯函数标准差关系曲线确定了最优的标准差;使用双阈值方法分割修复后的图像,结合孔洞填充与面积约束优化分割后的细胞形态。利用C2C12相差显微细胞图像数据集测试该算法,其中高细胞密度图像的分割精确率、召回率和F值分别为0.966 2、0.967 8和0.967 0,明显优于其他对比方法。实验结果表明,该方法在处理不同细胞密度的光照不均匀相差细胞图像时,均能实现光照均衡,且图像信息损失较小,分割计数结果准确。     

Asynchronous demodulation technique for use in vision sensor image classification and segmentation

This paper presents a novel technique for classification and segmentation of a multiple‐object image scene. Each object in the scene is tagged by a flashing LED operating at a specific frequency. The vision sensor, based on this technique, demodulates the captured light signal, omits the background illumination, and performs classification by assigning a unique ID‐tag to each region based on its flashing light frequency. The process is performed in‐pixel by an asynchronous demodulation and frequency identification circuit, which is designed in a standard 0.6µm CMOS technology. Simulation results confirm the validity of the proposed structure. At a frame rate of 250 fps the power consumption is 2.6µW/pixel, which is relatively low compared with those of sensors with similar functionality. The structure is intended for use as a low power, inexpensive solution in robot visual position feedback and localization. Copyright © 2009 John Wiley & Sons, Ltd.     

复杂背景条件下的电气设备图像实例分割算法

变电站巡检拍摄的电气设备可见光图像存在背景杂乱、目标轮廓不规则等特点,造成设备分割精度不高,影响智能巡检系统设备识别效果。基于此,提出一种改进的YOLACT++模型,实现设备目标精确实例分割。首先,设计了电气设备特征提取主干网络DAGNet,提升了网络对复杂背景下重要特征的关注度;同时在原型网络分支引入3D注意力模块SimAM,降低混乱背景对目标分割的干扰。使用某市8个区域58座110 kV变电站和86座35 kV变电站巡检所得避雷器、断路器等6类电气设备的1 730张可见光图像的标记数据集对该模型进行验证,实验结果表明,改进YOLACT++模型分割的AP_all指标为84.1%,相较原模型提高了4.4%,与YOLACT、Mask R-CNN和YOLOv8模型相比分别高出4.0%、9.3%、1.6%,较好地实现了6类电气设备的识别,可满足电力巡检中准确性和快速性的要求。     

基于门控金字塔融合的遥感影像海陆分割方法

海陆分割是通过遥感影像进行海岸线变化分析、资源管理等应用的重要基础,由于遥感影像场景复杂、陆地大小形状分布不均,海陆分割面临着误分类和边界分割不清等问题。针对上述问题,提出了一种用于遥感影像海陆分割的门控金字塔融合网络。首先通过基于注意力诱导的跨层聚合模块聚合两个深层特征,捕获全局上下文,准确而粗略地获取陆地的大小和形状信息。然后将聚合的全局特征送入门控融合模块,以全局信息为指导,在多尺度特征中选择有用的上下文信息,逐层优化边界细节并突出整个陆地区域。最后对每个侧输出进行全局监督。选取两组不同数据源的遥感影像进行实验,准确率分别为9913%和9898%,F1分数分别为9903%和9889%,mIoU分别为9826%和9797%。实验结果表明,与其他算法相比,该算法具有更好的分割效果。     

Development and testing of a deep learning-based strategy for scar segmentation on CMR-LGE images

Magnetic Resonance Materials in Physics, Biology and Medicine - The aim of this paper is to investigate the use of fully convolutional neural networks (FCNNs) to segment scar tissue in the left...     

水工混凝土裂缝像素级形态分割与特征量化方法

混凝土开裂问题在水工建筑物主体结构中普遍存在,裂缝检测一直是水工混凝土结构安全鉴定的重要内容.数字图像处理技术因具有效率高、成本低等优势而被广泛应用于结构表面裂缝检测中,形态分割与特征量化是其核心任务.针对传统图像处理人工干预较多、泛化能力较弱等不足,本文提出了一种基于深度卷积神经网络的水工混凝土裂缝像素级形态分割与特...     

面向图像语义分类的视觉单词集成学习方法

针对图像的特性,提出了1种视觉单词集成学习方法.该方法建立在3种初始映射方法的基础上,并充分利用图像的矩、纹理直方图、图像傅里叶描述子等图像视觉特征来分类图像语义.相对于3种初始映射方法,采用Boosting集成学习方法生成的视觉单词集合在图像语义分类上比单独使用任意1种映射方法生成的视觉单词集合有明显的提高.实验结果表明,本文所提出的视觉单词Boosting集成学习方法在图像语义区分性和描述能力方面是有效的,能充分反映人对图像内容的理解,具有很好的应用价值.     

Evaluation of five protocols for quantification of two-dimensional MR spectroscopic images

Five protocols were tested for quantification onin vivo two-dimensional spectroscopic imaging datasets. The datasets are duplo measurements from three individuals. The peaks forN-acetyl aspartate, creatines, and cholines were quantified. The first of the protocols is conventional integration of a selected interval around the peak. The others are various implementations of a Gauss-Newton-based least-squares time-domain fitting algorithm. Fitting a half-echo (free-induction decay curve, FID) to a full echo was compared; this was combined with or without filtering out the water signal using the Lanczos-Hankel singular-value decomposition (LHSVD). It appears that conventional integration yields consistent and accurate results in comparison to the fitting methods. The combination of LHSVD and echo fitting performs equally well; advantages and disadvantages are discussed. The protocols using FID fitting perform poorly with these experimental datasets.     

Quantitative analysis of GD-DTPA enhanced dynamic MR images by simplex minimization

The introduction of fast clinical imaging techniques enables the dynamic sampling of contrast agent kinetics. The quantity of data generated by such techniques demands the application of automated analysis tools to provide objective assessments of contrast washin/washout characteristics across an entire image. Using a simple compartmental model, the washin/washout characteristics may be defined in terms of a biexponential function. Curve fitting to this nonlinear function may be approached using a nonlinear least squares technique. We demonstrate an alternative technique, simplex minimization, to be more robust in a comparative study using Monte Carlo simulation. The application of the technique is illustrated by pixel-by-pixel analysis of dynamic data obtained from a patient with primary carcinoma of the breast.     

Real-time removal of impulse noise from MR images for radiosurgery applications

In the recent years, image processing techniques are used as a tool to improve detection and diagnostic capabilities in the medical applications. Among these techniques, medical image enhancement algorithms play an essential role in the removal of the noise, which can be produced by medical instruments and during image transfer. Impulse noise is a major type of noise, which is produced by medical imaging systems, such as MRI, computed tomography (CT), and angiography instruments. An embeddable hardware module, which can denoise medical images before and during surgical operations, could be very helpful. In this paper, an accurate algorithm is proposed for real-time removal of impulse noise in medical images. Our algorithm categorizes all image blocks into three types of edge, smooth, and disordered areas. A different reconstruction method is applied to each category of blocks for noise removal. The proposed method is tested on MR images. Simulation results show acceptable denoising accuracy for various levels of noise. Also, an field programmable gate array (FPGA) implementation of our denoising algorithm shows acceptable hardware resource utilization. Hence, the algorithm is suitable for embedding in medical hardware instruments such as radiosurgery devices.     

Tissue segmentation: a crucial tool for quantitative MRI and visualization of anatomical structures

Automatic or semi-automatic segmentation of tissue types or organs is well established for X-ray-based computed tomography, with its fixed grey-scale and tissue classes with well-established ranges of Hounsfield units. MRI is much more powerful with regard to soft tissue contrast and quantitative assessment of tissue properties (e.g., perfusion, diffusion, fat content), but the principle of signal generation and recording in MRI leads to inherent problems if simple threshold based segmentation procedures are applied. In this editorial in the special issue of MAGMA on tissue segmentation, a number of relevant methodical, scientific, and clinical aspects of reliable tissue segmentation using data recording by MRI are reported and discussed.