目的:采用计算机辅助定量分析腕管综合征(CTS)的灰阶超声图像,探讨其在定量分析CTS中的应用价值.方法:搜集60例CTS患者(77个异常手腕)和30例正常志愿者(60个正常手腕),其中17例为双侧卡压,在豌豆骨水平保留正中神经图像,共得到137组图像,并且在二维图像上对正中神经进行勾勒,提取正中神经卡压的客观数据,为CTS的诊断提供依据.结果:CTS患者病灶区域像素的均值、标准差、变异系数、偏度、峰度均比正常组大;CTS患者直方图熵较正常组大;正常组亮度熵较CTS组大;而CTS患者整个灰阶区域像素中的均值、中值较正常组低;CTS患者所有表示对比度的参数均高于正常组;CTS患者病灶的厚度、长度、面积、长轴、短轴、周长等参数均较正常组大.表明在CTS患者中,图像分布较正常对照组欠均匀,且CTS患者正中神经较正常肿胀,横截面积增大.结论:计算机辅助定量分析在CTS中表现良好,可定量分析正中神经卡压时图像的均匀性和对比度.

超声成像作为常见的医学成像技术,在诊断和治疗疾病方面发挥着重要作用.人工判读超声图像依赖于医生主观经验知识,其结果存在观察者间和观察者内的差异,并且过程耗时耗力,难以满足快速、批量的临床诊断需求.随着人工智能的快速发展,基于深度学习的图像分割模型被广泛地应用于医学图像分割领域,并取得了不俗成果.基于U-Net深度神经网络,对其结构进行优化改进,构建适用于超声图像正中神经分割的卷积神经网络模型.在网络中添加了规范层和丢弃层,下采样的最大池化层替换成卷积层;同时,采用指数线性单元激活函数,使用循环学习率,以增加网络模型的鲁棒性.测试结果表明,改进的U-Net模型在超声正中神经图像的自动分割方面表现良好,横切面、纵切面的Dice系数分别达到了78％与89％.该方法有望用于临床超声神经图像的辅助分析.

目的:运用改进U-Net深度网络学习定量评价腕管综合征正中神经的超声图像,确定基于改进U-Net深度网络学习的卷积神经网络模型,探讨其在定量评估腕管综合征正中神经卡压中的应用价值.方法:搜集213例经肌电图确诊的腕管综合征正中神经卡压患者及104例健康志愿者,213例正中神经卡压患者中60例为双侧卡压.对317例受检者行超声检查,在腕管处保存超声图像,共得到正中神经图像377组.由擅长肌骨超声的医师对377组图像进行勾勒.应用基于改进U-net深度网络学习的卷积神经网络模型,分割腕管综合征卡压的正中神经超声图像,定量分析提取横切以及纵切的正中神经超声图像的影像组学量化特征.结果:改进的U-Net深度网络可以很好地识别切割正中神经;改进的U-Net深度网络可以定量表示CTS中卡压的正中神经回声减低,区域明暗参数A、明暗参数I、对比明暗参数RI以及纹理参数Homo、纹理不均匀参数Cont差异均有统计学意义(P=0.000).结论:改进的U-Net模型在超声正中神经图像自动分割方面表现良好,可以定量分析腕管综合征正中神经卡压时灰度以及神经纹理均匀性.

淋巴结病变的诊断对于患者的治疗具有重要意义.在淋巴结病变的临床超声诊断中,通常只使用单一模态的B型超声图像.有时会采集B型超声和弹性超声或者采集B型超声和超声造影(CEUS)的双模态图像,很少情况下会采集全部三个模态图像.为了提高B型单模态的诊断性能,提出一种基于特权信息学习的淋巴结病变计算机辅助诊断(CAD)方法,在训练阶段使用三个模态图像,在测试阶段只使用B型.分别提取B型、弹性超声和CEUS图像的量化特征;在CAD模型中,训练样本为B型、弹性超声和CEUS多模态数据,测试样本只有B型;通过训练样本学习,得到结合特权信息的支持向量机(SVM+)模型,使用该模型对测试样本进行分类.试验结果表明,该方法的分类准确率、精度、敏感性、特异性和约登指数达到0.85、0.93、0.88、0.77和0.65,相较单模态B型超声训练的CAD模型,其分类结果分别提升了0.08、0.02、0.08、0.08和0.16.基于特权信息学习,提高了诊断精度,提升了计算机辅助诊断的性能.

课程项目是提高学生动手实践能力的重要教学手段.以数字图像处理课为例,探讨课程项目中的学习动机、策略与成绩间的关系.通过改进的学习动机策略量表对上海大学的292名学生进行问卷调查,有效样本为245人,计算学习动机、策略各分量及成绩间的相关系数,并根据课程项目成绩将学生分为高分组与低分组,比较两组间的差异.结果 表明:总体学习动机与总体学习策略间存在显著相关性(r=0.49,p＜0.001);学习动机的价值分量及总体学习策略均与课程项目成绩显著相关(p＜0.05),且在项目高低分两组间存在显著差异(p＜0.05);资源管理策略分量在项目高低分组间存在显著差异(p＜0.05).课程项目中学习动机与策略更强的学生能取得更好的项目成绩.

Noninvasive diagnosis of prostate cancer is of great importance for the treatment of patients.B-mode ultrasound and elastography are currently used in noninvasive diagnosis of prostate cancer.A technology for computer-aided diagnosis of prostate cancer is proposed based on dual-modal ultrasound namely elastography and B-mode ultrasound.Firstly,the quantitative features of B-mode images and elastographic images is extracted,including the characteristics of gray-level co-occurrence matrices, first-order statistical features,binary image features and regional features.Secondly,the canonical correlation analysis is used to fuse B-mode and elastographic features.Finally,the support vector machine is used for prostate disease classification.The experimental results on 313 prostate dual-modal images from 103 patients with prostatic diseases(47 malignant and 56 benign) show that the quantitative features are significant different between malignancy and benignancy.The sensitivity,specificity and accuracy of the classification are 78.7%,85.7% and 82.5%.This method is expected to be used for clinical noninvasive diagnosis of prostate cancer.

Objective To evaluate the diagnostic value of scoring method of combined conventional and contrast enhanced ultrasonography in the diagnosis of cervical metastatic lymph nodes.Methods Eight-two patients with 82 cervical lymph nodes were enrolled in the group.The scores on the conventional ultrasound were calculated from four aspects:the thickness of the the lymph nodes,the ratio of the long axis and the short axis of lymph node,the existence of lymph node hilus and the uniformity of lymph node.The scores on the contrast-enhanced ultrasound (CEUS)were calculated from the following aspects:the position of the enhancement in the arterial phase,the CEUS perfusion patterns and the intensity and uniformity of contrast enhancement in lymph nodes.Each feature was scored 1 point.For each lymph node,the combined scores was 0 to 8 points.Results Of the 82 enlarged lymph nodes,Forty-seven cervical metastatic lymph nodes and 35 reactive hyperplastic lymph node were analyzed.The scoring method that combination of conventional and contrast enhanced ultrasonography is the most valuable with 5.5 as the optimal diagnostic points.And the sensitivity and the specificity were 93.6％ and 82.9％ respectively.Conclusions The scoring method that combined conventional and contrast-enhanced ultrasonography was a semi-quantitative method and will be usefully in the differential diagnosis of cervical metastatic and reactive hyperplastic lymph node.

This paper discusses the course construction of Medical Ultrasound Technique in the biomedical engi-neering major in our university. Based on the purpose of talent training and corresponding course system of our bio-medical engineering major,the course content is constructed,and the teaching and practice methods are discussed. Specifically,for the training purpose of improving student's practice and self-study ability,the practical project of this course is detailedly introduced.

针对乳腺肿瘤超声图像分割,提出一种改进的反应扩散(RD)水平集分割算法。先使用Gabor各向异性扩散模型进行滤波,由此构造边界停止函数;再将该函数融入RD水平集演化方程,以控制曲线的演化得到乳腺肿瘤的边界。采用该方法和传统RD方法对77例病人的111幅乳腺超声图像进行分割实验,分割准确率分别为98.5%和98.0%,真阳性率分别为88.2%和82.7%,与金标准之间的均方根误差分别为3.6和4.6像素。结果表明,该改进算法可获得更加准确的乳腺肿瘤分割结果。

斑点噪声干扰超声图像的解读与疾病诊断.本文提出一种结合非局部均值(NLM)与伽柏各向异性扩散(GAD)的超声图像斑点降噪新方法.首先对图像进行GAD迭代自适应滤波,通过鲁棒的伽柏边缘检测算子有效区分边缘与噪声.接着进行NLM运算,针对某一像素点,其NLM估计值为GAD滤波后所有像素点的加权平均,而权重则由原始图像中该像素点所在区块与其余像素点所在区块间的相似度决定.由此该方法在NLM运算中兼顾了GAD滤波前后的信息以提高降噪性能.运用本文方法及七种传统方法对仿真图像与临床图像进行滤波实验.仿真实验表明,当噪声方差强达0.14时,本文方法较传统方法将品质因数、峰值信噪比与平均结构相似性提高13.29％、3.07％与0.88％.临床图像实验表明,无论在噪声去除还是在细节保留上,本文方法均优于传统方法.

区分淋巴结病变的良恶性具有积极临床意义。超声造影通过向血液中注射造影剂以动态显示组织中的新生血管及其血流灌注,是诊断淋巴结病变的新兴方法。针对病变淋巴结,提出一种从淋巴结超声造影图像中提取量化特征的方法,包括心动周期提取和子序列选择、淋巴结分割、纹理特征提取、统计学检验。对29个病人的41个淋巴结病灶的实验结果表明,提取的9个特征在良恶性淋巴结间存在显著性差异(P<0.05),有助于鉴别良恶性淋巴结。

Contrast-enhanced ultrasound (CEUS) is of great value for the diagnosis and treatment of vascular diseases. Extraction of carotid arterial contours is important for the measurement of morphological and elastic properties of arteries. Since manually tracing of arterial contours is time-consuming, subjective, and unrepeatable, computer-aided meth-ods are required. However, speckle noise in the CEUS images causes poor robustness and di?cult initialization in traditional computer-aided image segmentation methods. This paper integrates multi-scale fuzzy C-means clustering with particle swarm optimization to extract coarse boundaries of carotid arteries. Then boundaries are used as initial contours of the directional gradient vector flow (DGVF) model, and deform them until convergence to get final refined contours. Experimental results on 48 CEUS images from 14 patients show that the proposed method is superior to the traditional method, and can automati-cally and accurately extract boundaries of carotid arteries in CEUS images.