首页 | 本学科首页   官方微博 | 高级检索  
相似文献
 共查询到19条相似文献,搜索用时 140 毫秒
1.
目的本文结合深度学习特征(DF)和传统图像特征(HCF)特点,利用多分类器融合的方法建立一个乳腺肿瘤分类模型,并 深入评估和分析不同深度学习网络特征的肿瘤分类性能。方法回顾性分析106例乳腺肿瘤患者的头尾位和内外倾斜位投影 的全数字乳腺成像数据。首先从肿瘤区域提取23维HCF(12维形态及11维纹理特征),用t检验进行显著性特征选择;然后分别 从3 个卷积神经网络模型提取不同维度DF,在实验中,3 个不同深度学习网络产生了相应DF,分别是AlexNet,VGG16 和 GoogLeNet;最后结合2个投影数据的DF和HCF,采用多分类器的融合模型对特征进行训练和测试,实验重点分析不同DF在 肿瘤分类上的性能。结果结合DF和HCF建立的分类模型比使用单独HCF的分类模型表现出更好的性能;相比于其它网络框 架,DFAlexNet和HCF的结合表现出更高精度的分类结果。结论结合DF和HCF的特征方法建立一个分类模型,对于良恶性乳腺 肿瘤具有优秀的鉴别能力,且泛化能力更强,能作为临床辅助诊断工具。  相似文献   

2.
目的建立新的机器学习模型,从蛋白质数据集或全基因组蛋白质序列中预测外膜蛋白。方法采用分组重量编码和氨基酸组成计算蛋白质序列特征,采用F-score方法反向选择特征,采用支持向量机算法建立分类模型,在1 087条蛋白质序列构成的数据集上进行测试,评价预测模型的敏感性、特异性和预测精度,在多个细菌的全基因组蛋白质中预测外膜蛋白。结果新的序列组合编码方法与SVM相结合,区分外膜蛋白和α螺旋跨膜蛋白、球状蛋白、非外膜蛋白的准确度分别达到94.7%、96.4%和94.6%,经特征选择之后,分类准确度分别提高到95.7%、96.9%和95.9%,且在基因组数据集中的预测结果与已知事实相符度高。结论该方法预测准确度优于其他基于序列特征的预测方法,可用于在基因组序列中预测和筛选新的外膜蛋白。  相似文献   

3.
目的:评估亨廷顿病(HD)患者运动的特殊类型,特别是运动功能减退(即运动幅度的减少)和运动徐缓(即运动缓慢、困难)在步态障碍中的各自作用。方法:采用视频运动分析系统记录15例早期HD患者的运动学、空间(步长、速度)、时间(步调、速度和跨步时间)和步态角度参数(踝关节活动范围),并将之与15例对照者和15例帕金森病(PD)患者进行比较。采用空间(步长减少)和步态角度参数(踝关节活动范围减少)对运动功能减退进行研究,而运动功能亢进以踝关节活动范围增加为特征,同时采用时间参数(步调、跨步时间)对运动徐缓(步行速度降低)进行评估。探讨临床症状(运动功能障碍、舞蹈症、全身残疾和认知功能障碍)和CAG重复数对步态障碍的影响。结果:HD患者的步行速度和步调减慢,而跨步时间延长(即运动徐缓),且具有明显的个体差异,PD患者的步调正常。尽管HD患者的步长以运动功能减退样步态(如同PD患者的表现)为特征,但尚无证据表明,其步长显著减少。踝关节活动角度分析显示,HD患者同时出现运动功能亢进和运动功能减退,导致步态异常。HD患者的步行速度与UHDRS的运动部分有关。结论:HD患者的步态主要以定时障碍为特征:步态的时间参数表现为运动徐缓,并且个体内的严重程度存在差异。  相似文献   

4.
随着计算机技术的飞速发展以及人机交互技术的广泛应用,基于视频的表情识别逐渐成为研究热点之一,并逐渐实用化。本文提出了一种基于视频的情感时空融合特征提取算法,并用于表情识别。首先获取情感视频的时空特征点和其对应的立方体(cuobids),然后融合Piotr Dollar提出的描述算子和CBP_TOP描述算子所提取的cuobids的特征向量作为时空特征点最终的特征向量,最后采用“词袋模型”方法来提取情感视频最终的表情特征,并用于后续的表情分类。仿真实验表明此算法在保证识别精度的基础上大大提高了识别速率。  相似文献   

5.
针对医学步态分析中的运动目标检测问题,提出基于最小错误率的贝叶斯决策规则的方法。该方法由变化检测、变化分类、前景目标提取和背景更新四部分组成。变化检测采用自适应阈值法来二值化变化点和非变化点。变化分类基于颜色共生特征向量,采用贝叶斯规则进行决策,前景对象的提取融合了时间差分法和减背景法。针对复杂场景中背景的“渐变”和“突变”情况,提出不同的背景更新策略。实验表明,该方法和包含有摇动的树枝或者灯的开关等复杂背景中能准确地提取运动目标,因此可用在医学步态分析的研究中。  相似文献   

6.
目的 探讨基于MRI平扫构建的影像组学模型用于鉴别诊断软骨肉瘤与内生软骨瘤的价值。方法 回顾性分析68例软骨源性肿瘤(软骨肉瘤27例,内生软骨瘤41例),将其随机分配到训练组(n=46)与验证组(n=22)。首先由2名放射科医师独立提取平扫T1WI和T2WI-FS序列中肿瘤所有层面的影像组学特征,采用组内相关系数(ICC)评价2名医师提取组学特征的一致性;然后使用方差选择法、单变量特征选择、最小绝对收缩与选择算子算法(LASSO)对组学特征进行筛选和降维,使用多因素逻辑回归分析构建基于T1WI和T2WI-FS序列的组学模型,采用受试者工作特征曲线(ROC)评估组学模型的诊断效能, 并与放射科医师采用常规MR序列的诊断效能进行对比。结果 2名放射科医师独立提取患者T1WI和T2WI-FS序列影像组学特征的一致性良好(ICC值范围为0.779~0.923)。在T1WI序列筛选出10个组学特征,在T2WI-FS序列筛选出11个组学特征,两个序列的组 学模型在训练组中AUC分别为0.990和0.925;在验证组中AUC分别0.915和0.855,模型之间的诊断效能差异均无统计学意义(P>0.05)。在所有病例中,T1WI、T2WI-FS序列组学模型与常规MRI诊断的AUC分别为0.955、0.901、0.569,基于两个序列的组学模型诊断准确性均高于放射科医生的诊断效能(P<0.001)。结论 基于MRI平扫T1WI和T2WI-FS序列构建的影像组学模型能用于鉴别诊断软骨肉瘤与内生软骨瘤。  相似文献   

7.
目的 针对深度学习在舌象分类中训练数据量大、训练设备要求高、训练时间长等问题,提出一种基于迁移学习的全连接神经网络小样本舌象分类方法。方法 应用经ImageNet海量数据集训练后的卷积Inception_v3网络提取舌象点、线等有效特征,再使用全连接神经网络对特征进行训练分类,将深度学习网络学习到的图像知识迁移到舌象识别任务中。利用舌象数据集进行训练、测试。结果 与典型舌象分类方法K最近邻(KNN)算法、支持向量机(SVM)算法和卷积神经网络(CNN)深度学习方法相比,本实验使用的两种方法(Inception_v3+2NN和Inception_v3+3NN)具有较高的舌象分类识别率,准确率分别达90.30%和93.98%,且样本训练时间明显缩短。结论 与KNN算法、SVM算法和CNN深度学习方法相比,基于迁移学习的全连接神经网络舌象分类方法可有效提高舌象分类的准确率、缩短网络的训练时间。  相似文献   

8.
针对医学步态分析中的运动目标检测问题,提出了基于最小错误率的贝叶斯决策规则的方法.该方法由变化检测、变化分类、前景目标提取和背景更新四部分组成.变化检测采用自适应阈值法检测二值化变化点和非变化点.变化分类基于颜色共生特征向量,采用贝叶斯规则进行决策,前景对象的提取融合了时间差分法和减背景法.针对复杂场景中背景的"渐变"和"突变"情况,提出了不同的背景更新策略.实验表明,该方法能将包含有摇动的树枝或者灯的开关等复杂背景中运动目标准确地提取,可用在医学步态分析的研究中.  相似文献   

9.
目的:研究如何整合并优化影像、神经认知评价和生物标志测量等多来源多模态数据以提高阿尔兹海默症(Alzheimer disease,AD)发展阶段和转化的分类预测准确率。方法:基于阿尔兹海默症影像计划(Alzheimer’s disease neuroimaging initiative,ADNI)2004—2018年4个阶段的样本数据,包括从核磁共振成像(magnetic resonance imaging,MRI)影像数据提取的脑图像特征数据、神经认知量表(简易精神状态测量量表和ADAS?Cog13量表)数据、生物标志测量数据(Abeta、Tau和p?Tau蛋白及ApoE4基因型)。基于783个样本的基线数据建立二分类和多分类Logistic回归模型用于疾病发展阶段的两两和同时分类预测。基于具有轻度认知障碍(mild cognitive impairment,MCI)状态的352个样本的纵向数据建立二分类Logistic回归模型并用于转化状态的分类预测。将脑图像特征变量、量表数据和生物标志加入到基准模型中,通过交叉验证方法随机划分数据集,并计算准确率、查准率、召回率、F1得分和ROC曲线下面积(area under curve,AUC)等指标进行综合比较,得到最优多模态组合的分类预测模型。结果:对于AD发展阶段的分类,结合脑图像特征数据、量表数据和生物标志数据建立二分类Logistic模型表现最佳,区分AD组和正常组、MCI组和正常组以及AD组和MCI组的准确率分别达到了100.00%、77.18%和89.58%;AUC值分别为100.00%、85.52%和96.39%,比仅用脑图像数据进行进展阶段的分类预测有显著提高。对于MCI是否转化的分类预测,脑图像特征数据结合量表数据和生物标志能最大限度地提高准确率,从86.69%提高到90%以上;相应的AUC值从89.21%提高到94.06%。结论:结合多来源数据能提高AD疾病进展阶段和转化的分类预测准确率,为临床诊断AD所处的发展阶段和转化提供理论上的支持。  相似文献   

10.
医学步态分析中的复杂场景下运动目标检测技术   总被引:1,自引:0,他引:1  
针对医学步态分析中的复杂场景下运动目标检测问题,提出了基于贝叶斯决策规则的方法.该方法由变化检测、变化分类、前景目标提取和背景更新四部分组成.变化检测采用自适应阈值法来二值化变化点和非变化点,变化分类基于颜色共生特征向量,采用贝叶斯规则进行决策,前景对象的提取融合了时间差分法和减背景法.针对复杂场景中背景的"渐变"和"突变"情况,提出了不同的背景更新策略.实验表明,该方法在包含有摇动的树枝,或者灯的开关等复杂背景中能准确地提取运动目标,因此可用在医学步态分析的研究中.  相似文献   

11.
Feature extraction and classification are considered to be the major tasks in image processing applications. This paper proposes a novel method to extract the features of a color image for classification. The proposed method, Dominant Local Texture-Color Patterns (DLTCP) is based on the Dominant Texture and Dominant Color channels in a RGB color space. The dominant texture pattern represents a channel among RGB with maximum variations in the texture and the dominant color pattern represents the color channel with the maximum pixel intensity. The combination of channels with dominant texture pattern and dominant color pattern is assigned a unique value which is used to extract the features of an image. The proposed texture-color features is tested for rotational, illumination and scale invariance property using the color images taken from Outex and Vistex databases. It is experimentally shown that the proposed method achieves the highest accuracy in classification using K-Nearest Neighbor (KNN) classifier under various challenges.  相似文献   

12.
A portable and wireless activity monitoring system was developed for the estimation of temporal gait parameters. The new system was built using three-axis accelerometers to automatically detect walking steps with various walking speeds. The accuracy of walking step-peak detection algorithm was assessed by using a running machine with variable speeds. To assess the consistency of gait parameter analysis system, estimated parameters, such as heel-contact and toe-off time based on accelerometers and footswitches were compared for consecutive 20 steps from 19 individual healthy subjects. Accelerometers and footswitches had high consistency in the temporal gait parameters. The stance, swing, single-limb support, and double-limb support time of gait cycle revealed ICCs values of 0.95, 0.93, 0.86, and 0.75 on the right and 0.96, 0.86, 0.93, 0.84 on the left, respectively. And the walking step-peak detection accuracy was 99.15% (±0.007) for the proposed method compared to 87.48% (±0.033) for a pedometer. Therefore, the proposed activity monitoring system proved to be a reliable and useful tool for identification of temporal gait parameters and walking pattern classification.  相似文献   

13.
目的 提出一种并行神经网络分类方法,以提高对正常搏动、室上性异位搏动、心室异位搏动、融合搏动4种心律失常的分类性能。方法 首先进行心电信号去噪、小尺度心拍和大尺度心拍的分割、数据增强等预处理;然后基于深度学习理论,应用密集连接卷积神经网络改善人工提取波形特征的局限性,并结合双向长短时记忆网络和高效通道注意力网络,以增强提取波形时序特征和重要特征的功能;接着采用并行网络结构,同时输入小尺度心拍和大尺度心拍的的波形特征,以提高心律失常分类的准确性;最后使用Softmax函数实现对心律失常的4分类任务。结果 利用MIT-BIH心律失常数据库和3组实验验证所提方法。多种并行网络模型分类性能对比实验和不同心拍输入方式下,各分类模型性能对比实验得出所提分类模型的总体准确率、平均灵敏度和平均特异性分别达到99.36%、96.08%、99.41%;并行网络分类模型收敛性能分析实验得出分类模型每次训练时间为41 s。结论 并行多网络分类方法在保持较高总体准确率的同时,平均灵敏度、平均特异性以及训练时间均有改善,该方法有望为心律失常临床诊断提供新的技术方案。  相似文献   

14.
通过采集腿部肌肉5个通道的肌音信号,利用3层决策树对跑步、上楼、下楼、走路、静止5种步态动作进行模式识别研究。在决策树的第1层和第2层,应用双阈值门限法识别静止和跑步两种步态模式,在第3层,提出基于步态信号的自适应不等长分割算法以及改进的模糊熵算法,利用线性分类器对走路、上楼、下楼进行分类识别。结果表明:双门限阈值法可有效地对静止和跑步进行识别,当采用改进的模糊熵特征时,对走路、上楼、下楼3种步态模式的分类准确率达到了94.87%;而当综合利用近似熵、样本熵和改进的模糊熵3种特征时,其分类准确率达到了98.76%。  相似文献   

15.
目的 探讨基于支持向量机(SVM)构建的人工智能辅助诊断模型对椎弓根螺钉钉道完整性进行超声鉴别与验证的方法研究。 方法 本文提出了一种基于超声图像智能分析的椎弓根钉道完整性评估方法。数据采自4个新鲜人体胸腰椎标本。预建立钉道50个,共800张超声图像(其中钉道完整与破损的样本各400个),采用五折交叉验证的方法对样本进行训练集与测试集的划分,对人工智能辅助诊断模型进行训练及测试。首先对超声图像进行裁剪,并采用亮度映射方法进行图像增强得到易于计算机判断识别且排除无效信息干扰的超声图像;然后通过灰度共生矩阵算法进行纹理特征提取,并采用支持向量机模型对正常和严重破损样本的初始分类模型进行搭建;其次,使用灰度分布得到用于区分前景和背景的阈值,并通过设计的损失函数得到得到钉道同心圆的半径;最后将同心圆外部图像的熵、方差、对比度、能量、平均绝对偏差作为第二类特征,最后进行轻微破损样本和未破损样本的二次分类模型搭建。 结果 初始分类的准确率为74.75%,特异性为71.81%,灵敏度为81.5%,F1值为76.35%,假正率为32%,假负率为18.5%。二次分类的准确率为93.75%,特异性为91.55%,灵敏度为97.5%,F1值为94.43%,假正率为9%,假负率为2.5%。二次类准确率与初始分类相比较,准确率提升19%。 结论 本文提出的基于SVM机器学习模型的方法可较为准确地检测椎弓根钉道的破损情况,且准确率较高,可用于术中实时判断椎弓根钉道的状态。  相似文献   

16.
This paper focuses on the issue of extracting retina vessels with supervised approach. Since the green channel in the retina image has the best contrast between vessel and non-vessel, this channel is used to separate vessels. In our approach we are proposing a technique of using gray-level co-occurrence matrix method for composition of the retinal images. It is based on fact that the co-occurrence matrix of retina image describes the transition of intensities between neighbour pixels, indicating spatial structural information of retina image. So, we first extract the features vector based on specified characteristics of the gray-level co-occurrence matrix and then we use these features vector to train a neural network approach for the classification method which makes our proposed approach more effective. Obtained results from the experiments in DRIVE and STARE database shows the advantage of the proposed method in contrast to current methods. This advantage is evaluated by the criteria of sensitivity, specificity, area under ROC and accuracy. The result of such a conversion as the input vector of a multilayer perceptron neural network will be trained and tested. Although in recent years different methods have been presented in this respect, but results of simulation shows that the proposed algorithm has a very high efficiency than the other researches.  相似文献   

17.
Multiscale entropy (MSE) is one of the popular techniques to calculate and describe the complexity of the physiological signal. Many studies use this approach to detect changes in the physiological conditions in the human body. However, MSE results are easily affected by noise and trends, leading to incorrect estimation of MSE values. In this paper, singular value decomposition (SVD) is adopted to replace MSE to extract the features of physiological signals, and adopt the support vector machine (SVM) to classify the different physiological states. A test data set based on the PhysioNet website was used, and the classification results showed that using SVD to extract features of the physiological signal could attain a classification accuracy rate of 89.157%, which is higher than that using the MSE value (71.084%). The results show the proposed analysis procedure is effective and appropriate for distinguishing different physiological states. This promising result could be used as a reference for doctors in diagnosis of congestive heart failure (CHF) disease.  相似文献   

18.
This paper presents the new automated detection method for electrocardiogram (ECG) arrhythmias. The detection system is implemented with integration of complex valued feature extraction and classification parts. In feature extraction phase of proposed method, the feature values for each arrhythmia are extracted using complex discrete wavelet transform (CWT). The aim of using CWT is to compress data and to reduce training time of network without decreasing accuracy rate. Obtained complex valued features are used as input to the complex valued artificial neural network (CVANN) for classification of ECG arrhythmias. Ten types of the ECG arrhythmias used in this study were selected from MIT-BIH ECG Arrhythmias Database. Two different classification tasks were performed by the proposed method. In first classification task (CT-1), whether CWT-CVANN can distinguish ECG arrhythmia from normal sinus rhythm was examined one by one. For this purpose, nine classifiers were improved and executed in CT-1. Second classification task (CT-2) was to recognize ten different ECG arrhythmias by one complex valued classifier with ten outputs. Training and test sets were formed by mixing the arrhythmias in a certain order. Accuracy rates were obtained as 99.8% (averaged) and 99.2% for the first and second classification tasks, respectively. All arrhythmias in training and test phases were classified correctly for both of the classification tasks.  相似文献   

19.
The classification of heartbeats is crucial to identify an arrhythmia. This paper proposes a new method that combines independent component analysis (ICA) with sparse representation-based classification (SRC) to distinguish eight types of heartbeats. We use ICA to extract useful features from heartbeats. A feature vector consists of 100 ICA features along with a RR interval. We use SRC to compute a sparse representation of a test feature vector with respect to all training feature vectors. The type of a test feature vector is determined using the concentration degree of sparse coefficients on each heartbeat type. For experimental purposes, 9800 heartbeats are extracted from the MIT-BIH electrocardiogram (ECG) database. The results show that our proposed method performs better than conventional methods, with 98.35% accuracy and 94.49%–100% sensitivities to several heartbeat types.  相似文献   

设为首页 | 免责声明 | 关于勤云 | 加入收藏

Copyright©北京勤云科技发展有限公司  京ICP备09084417号