医学影像处理的深度学习可解释性研究进展

随着医学影像数据的迅速增长，传统的影像分析方法给医生带来巨大挑战。利用计算机视觉技术提供自动或半自动辅助诊断，可大大缓解人工阅片压力，提高诊断的准确性，促进医疗流程的标准化建设等。目前，深度学习卷积神经网络在医学影像处理中已取得不俗表现，但深度学习“黑匣子”的不可解释性阻碍了智能医疗诊断的发展。为增强对医学影像数据处理的深度学习可解释性的了解，对近几年相关研究进展进行了综述。首先，综述了深度学习在医学领域的应用现状及面临的问题，对神经网络的可解释性内涵进行了讨论；然后，从现有深度学习可解释性的常见方法出发，重点讨论了医学影像处理的深度学习可解释性研究进展；最后，探讨了医学影像处理的深度学习可解释性的发展趋势。     

基于快速神经网络架构搜索的鲁棒图像水印网络算法

为解决深度学习在图像水印算法中计算量大且模型冗余的问题，提高图像水印算法在抵抗噪声、旋转和剪裁等攻击时的鲁棒性，提出基于快速神经网络架构搜索（neural architecture search，NAS）的鲁棒图像水印网络算法。通过多项式分布学习快速神经网络架构搜索算法，在预设的搜索空间中搜索最优网络结构，进行图像水印的高效嵌入与鲁棒提取。首先，将子网络中线性连接的全卷积层设置为独立的神经单元结构，并参数化表示结构单元内节点的连接，预先设定结构单元内每个神经元操作的搜索空间；其次，在完成一个批次的数据集训练后，依据神经元操作中的被采样次数和平均损失函数值动态更新概率；最后，重新训练搜索完成的网络。水印网络模型的参数量较原始网络模型缩减了92%以上，大大缩短了模型训练时间。由于搜索得到的网络结构更为紧凑，本文算法具有较高的时间性能和较好的实验效果，在隐藏图像时，对空域信息的依赖比原始网络更少。对改进前后的2个网络进行了大量鲁棒性实验，对比发现，本文算法在CIFAR-10数据集上对抵抗椒盐噪声和旋转、移除像素行（列）等攻击优势显著；在ImageNet数据集上对抵抗椒盐高斯噪声、旋转、中值滤波、高斯滤波、JPEG压缩、裁剪等攻击优势显著，特别是对随机移除行（列）和椒盐噪声有较强的鲁棒性。     

Digital image correlation in experimental mechanics and image registration in computer vision: Similarities,differences and complements

Digital image correlation and image registration or matching are among the most widely used techniques in the fields of experimental mechanics and computer vision, respectively. Despite their applications in separate fields, both techniques primarily involve detecting the same physical points in two or more images. In this paper, a brief technical comparison of the two techniques is reviewed, and their similarities and differences as well as complements are presented. It is shown that some concepts from the image registration or matching technique can be applied to the digital image correlation technique to substantially enhance its performance, which can help broaden the applications of digital image correlation in scientific research and engineering practice.     

Evolution of downsized crescent-shaped dune in wind tunnel experiment

The migration of a downsized crescent-shaped dune was investigated in a wind tunnel experiment.Quantified upwind influx and vertical oscillation of the sand bed were introduced to modulate the saturation level of the sand flux above the dune surface to affect dune evolution.The evolution was recorded by top-view photography and then abstracted as the evolution of self-defined characteristic quantities using a digital image processing algorithm.The results showed that,in contrast to the case for spanwise quantities,the evolution of streamwise quantities corresponds to a linear increase in the modulation magnitude more positively and in a monotonic and convergent manner.In contrast with quantities on the windward face,the changes in quantities with respect to the horns were nonmonotonic with time and almost uncorrelated with the variation in modulation strength,which reveals the distinctiveness of leeside evolution.     

Rapid and label-free classification of pathogens based on light scattering,reduced power spectral features and support vector machine

The rapid identification of pathogens is crucial in controlling the food quality and safety. The proposed system for the rapid and label-free identification of pathogens is based on the principle of laser scattering from the bacterial microbes. The clinical prototype consists of three parts: the laser beam, photodetectors, and the data acquisition system. The bacterial testing sample was mixed with 10 mL distilled water and placed inside the machine chamber. When the bacterial microbes pass by the laser beam, the scattering of light occurs due to variation in size, shape, and morphology. Due to this reason, different types of pathogens show their unique light scattering patterns. The photo-detectors were arranged at the surroundings of the sample at different angles to collect the scattered light. The photodetectors convert the scattered light intensity into a voltage waveform. The waveform features were acquired by using the power spectral characteristics, and the dimensionality of extracted features was reduced by applying minimal-redundancy-maximal-relevance criterion (mRMR). A support vector machine (SVM) classifier was developed by training the selected power spectral features for the classification of three different bacterial microbes. The resulting average identification accuracies of E. faecalis, E. coli and S. aureus were 99%, 87%, and 94%, respectively. The overall experimental results yield a higher accuracy of 93.6%, indicating that the proposed device has the potential for label-free identification of pathogens with simplicity, rapidity, and cost-effectiveness.     

Experimental investigation of the effect of temperature on the strength of polyimide contact structures

A new photogrammetry based measurement technique for contact areas in line contact structures is developed. From the data collected by this technique, a method for measuring the contact strength of contact structures in high temperature applications is proposed. The contact strength of a line contact structure of polyimide at different temperatures is measured; the results show that the increase in temperature decreases the contact strength of the structure. The effect of temperature on the contact strength is studied by investigating the occurrence and evolution of yielding core within the structure at different temperatures. The decrease in the yield strength and elastic modulus of the polyimide material at high temperature induces core yielding at lower loads and eases the propagation of core yielding to the surface.     

Hitting probabilities and the Hausdorff dimension of the inverse images of a class of anisotropic random fields

Let X = {X(t):t ∈ R~N} be an anisotropic random field with values in R~d.Under certain conditions on X,we establish upper and lower bounds on the hitting probabilities of X in terms of respectively Hausdorff measure and Bessel-Riesz capacity.We also obtain the Hausdorff dimension of its inverse image,and the Hausdorff and packing dimensions of its level sets.These results are applicable to non-linear solutions of stochastic heat equations driven by a white in time and spatially homogeneous Gaussian noise and anisotropic Guassian random fields.     

Local zone wise elastic and plastic properties of electron beam welded Ti–6Al–4V alloy using digital image correlation technique: A comparative study between uniform stress and virtual fields method

Joining of materials using welding results in the formation of material zones with varying microstructure across the weld. Extraction of the mechanical properties of those individual heterogeneous zones are important in designing components and structures comprised of welds. In this study, the zone wise local extraction of the elastic and plastic properties of an electron beam welded Ti–6Al–4V titanium alloy has been carried out using both the uniform stress method (USM) and the virtual fields method (VFM) involving digital image correlation (DIC) technique. The surface strain field obtained using DIC technique from a transverse weld specimen tensile testing is used for extracting the zone wise strain evolution. Initially, using uniform stress assumption, zone wise full range stress–strain curves are extracted. In USM methodology, the elastic and plastic material models are fitted to the zone wise stress–strain curves and required parameters are extracted from it. But inherent disadvantage is lot of images need to be processed for the parameter extraction. Recently, VFM is gaining lot of popularity in characterization domain as it is robust, accurate and faster. VFM is based on the principle of virtual work where, the weak form of local equilibrium equations and kinematically admissible virtual displacement fields are utilized for parameter extraction. Hollomon׳s power law is used here as the hardening rule. Young׳s modulus, Poisson׳s ratio, yield stress, strength coefficient and strain hardening exponent are the parameters extracted zone wise using both USM and VFM. A Vicker׳s microhardness measurement is also conducted across the weld zone towards mapping the strength behavior. Fusion zone has reported higher yield strength, strength coefficient and Poisson׳s ratio. Young׳s modulus value is found decreasing from base metal towards the fusion zone. The trend observed in parameter variation across the weld zone obtained by both USM and VFM compares very well. Due to various advantages associated with VFM technique it is generally recommended for parameter extraction.     

A new class of conformable spectral observers for signal reconstruction

In this work, the design of spectral observers for signal reconstruction based on Kalman filters is performed and evaluated. The conformable derivative and the beta‐derivative were used to design the Kalman filters. Both derivatives satisfy the same formulas of the classical derivation, eg, the chain rule. The derivative order, the Ricatti equation parameters, and the observers tuning parameters were optimized using an optimization algorithm based on the bat's echolocation behavior (Bat algorithm). The simulation results showed the advantages of using the proposed observers for the signal reconstruction.