基于双模态深度自编码的孤立性肺结节诊断方法

近年来,深度学习技术在肺癌诊断方面得到了广泛的应用,但现有的研究主要集中于肺部CT图像。为了有效提高肺结节的诊断性能,提出一种基于双模态深度降噪自编码的肺结节诊断方法。首先,分别从肺部CT和PET图像中得到肺结节区域的特征信息;然后,以候选结节的PET/CT图像作为整个深度自编码网络的输入,并对高层信息进行学习;最后,采用融合策略对多种特征进行融合并将其作为整个框架的输出。实验结果表明,提出的方法可以达到92.81%的准确率、91.75%的敏感度和1.58%的特异性,且优于其他方法的诊断性能,更适用于肺结节良/恶性的辅助诊断。     

Medical Sign Recognition of Lung Nodules Based on Image Retrieval with Semantic Features and Supervised Hashing

Sign recognition is important for identifying benign and malignant nodules. This paper proposes a new sign recognition method based on image retrieval for lung nodules. First, we construct a deep learning framework to extract semantic features that can effectively represent sign information. Second, we translate the high-dimensional image features into compact binary codes with principal component analysis (PCA) and supervised hashing. Third, we retrieve similar lung nodule images with the presented adaptive-weighted similarity calculation method. Finally, we recognize nodule signs from the retrieval results, which can also provide decision support for diagnosis of lung lesions. The proposed method is validated on the publicly available databases: lung image database consortium and image database resource initiative (LIDC-IDRI) and lung computed tomography (CT) imaging signs (LISS). The experimental results demonstrate our retrieval method substantially improves retrieval performance compared with those using traditional Hamming distance, and the retrieval precision can achieve 87.29% when the length of hash code is 48 bits. The entire recognition rate on the basis of the retrieval results can achieve 93.52%. Moreover, our method is also effective for real-life diagnosis data.     

无监督混阶栈式稀疏自编码器的图像分类学习

目前多数图像分类的方法是采用监督学习或者半监督学习对图像进行降维,然而监督学习与半监督学习需要图像携带标签信息。针对无标签图像的降维及分类问题,提出采用混阶栈式稀疏自编码器对图像进行无监督降维来实现图像的分类学习。首先,构建一个具有三个隐藏层的串行栈式自编码器网络,对栈式自编码器的每一个隐藏层单独训练,将前一个隐藏层的输出作为后一个隐藏层的输入,对图像数据进行特征提取并实现对数据的降维。其次,将训练好的栈式自编码器的第一个隐藏层和第二个隐藏层的特征进行拼接融合,形成一个包含混阶特征的矩阵。最后,使用支持向量机对降维后的图像特征进行分类,并进行精度评价。在公开的四个图像数据集上将所提方法与七个对比算法进行对比实验,实验结果表明,所提方法能够对无标签图像进行特征提取,实现图像分类学习,减少分类时间,提高图像的分类精度。     

CT影像肺结节分割研究进展

准确分割肺结节在临床上具有重要意义。计算机断层扫描（computer tomography,CT）技术以其成像速度快、图像分辨率高等优点广泛应用于肺结节分割及功能评价中。为了进一步对肺部CT影像中的肺结节分割方法进行探索,本文对基于CT影像的肺结节分割方法研究进行综述。1）对传统的肺结节分割方法及其优缺点进行了归纳比较;2）重点介绍了包括深度学习、深度学习与传统方法相结合在内的肺结节分割方法;3）简单介绍了肺结节分割方法的常用评价指标,并结合部分方法的指标表现展望了肺结节分割方法研究领域的未来发展趋势。传统的肺结节分割方法各有优缺点和其适用的结节类型,深度学习分割方法因普适性好等优点成为该领域的研究热点。研究者们致力于如何提高分割结果的准确度、模型的鲁棒性及方法的普适性,为了实现此目的本文总结了各类方法的优缺点。基于CT影像的肺结节分割方法研究已经取得了不小的成就,但肺结节形状各异、密度不均匀,且部分结节与血管、胸膜等解剖结构粘连,给结节分割增加了困难,结节分割效果仍有很大提升空间。精度高、速度快的深度学习分割方法将会是研究者密切关注的方法,但该类方法仍需解决数据需求量大和网络模型超参数的确定等问题。     

改进U型残差网络用于肺结节检测

针对计算机断层扫描（CT）影像中肺结节检测灵敏度较低,且存在大量假阳性的问题,提出一种改进的U型残差网络用于肺结节检测。采取U-net网络的U型结构并利用残差学习方式构建深层次网络,同时引入自校正卷积增加特征的信息提取能力,进行通道间与局部信息增强,有利于检测不同形态的结节;通过引入的通道注意力机制,对特征提取过程中的特征进行重标定,实现自适应学习特征权重,进一步提高检测的准确率;引入DR loss作为该算法的分类损失函数,用于解决数据正负样本失衡问题。在LUNA16数据集对所提算法进行了验证,CPM得分达到0.901,提高了肺结节检测的灵敏度,而且有效降低了检测结果的平均假阳性个数,可有效辅助放射科医师对肺结节进行检测。     

融合深度网络和浅层纹理特征的甲状腺结节癌变超声图像诊断

目的 在甲状腺结节图像中对甲状腺结节进行良恶性分析,对于甲状腺癌的早期诊断有着重要的意义。随着医疗影像学的发展,大部分的早期甲状腺结节可以在超声图像中准确地检测出来,但对于结节的性质仍然缺乏准确的判断。因此,为实现更为准确的早期甲状腺结节良恶性超声图像诊断,避免不必要的针刺或其他病理活检手术、减轻病患生理痛苦和心理压力及其医疗费用,提出一种基于深度网络和浅层纹理特征融合的甲状腺结节良恶性分类新算法。方法 本文提出的甲状腺结节分类算法由4步组成。首先对超声图像进行尺度配准、人工标记以及图像复原去除以增强图像质量。然后,对增强的图像进行数据扩展,并作为训练集对预训练过的GoogLeNet卷积神经网络进行迁移学习以提取图像中的深度特征。同时,提取图像的旋转不变性局部二值模式（LBP）特征作为图像的纹理特征。最后,将深度特征与图像的纹理特征相融合并输入至代价敏感随机森林分类器中对图像进行良恶性分类。结果 本文方法在标准的甲状腺结节癌变数据集上对甲状腺结节图像取得了正确率99.15%,敏感性99.73%,特异性95.85%以及ROC曲线下面积0.997 0的的好成绩,优于现有的甲状腺结节图像分类方法。结论 实验结果表明,图像的深度特征可以描述医疗超声图像中病灶的整体感官特征,而浅层次纹理特征则可以描述超声图像的边缘、灰度分布等特征,将二者统一的融合特征则可以更为全面地描述图像中病灶区域与非病灶区域之间的差异以及不同病灶性质之间的差异。因此,本文方法可以准确地对甲状腺结节进行分类从而避免不必要手术、减轻病患痛苦和压力。     

Automatic classification for solitary pulmonary nodule in CT image by fractal analysis based on fractional Brownian motion model

Perfusion computed tomography (CT) method has been used to differentiate malignant pulmonary nodules from benign nodules based on the assessment for the change of the CT attenuation value within the pulmonary nodules. Instead of using the change of the CT attenuation value, a set of fractal features based on fractional Brownian motion model is proposed in this paper to automatically distinguish malignant nodules from benign nodules. In a set of 107 CT images from 107 different patients with each image containing a solitary pulmonary nodule, our experimental results obtained from a support vector machine classifier show that the accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and the area under the ROC curve are 83.11%, 90.92%, 71.70%, 80.05%, 87.52%, and 0.8437, respectively, by using the proposed fractal-based feature set. Such a result outperforms the conventional method of using the change of the CT attenuation value as the feature for classification. When combining this conventional method with our proposed fractal-based method, the accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and the area under the ROC curve can be promoted to 88.82%, 93.92%, 82.90%, 87.30%, 90.20%, and 0.9019, respectively. In other words, a high performance of pulmonary nodule classification can be achieved with a single post-contrast CT scan.     

基于3D CNN的肺结节假阳性筛查模型

通过肺部CT影像进行肺结节检测是肺癌早期筛查的重要手段,而候选结节的假阳性筛查是结节检测的关键部分.传统的结节检测方法严重依赖先验知识,流程繁琐,性能并不理想.在深度学习中,卷积神经网络可以在通用的学习过程中提取图像的特征.该文以密集神经网络为基础设计了一个三维结节假阳性筛查模型—三维卷积神经网络模型(TDN-CNN)...     

肺结节球表面网格向量化特征分类

目的 基于球谐函数与容斥映射算法向量化球面表面纹理与结节形状用以进行胸部CT图像肺结节良恶性判定。区别于基于深度学习解决肺结节良恶性筛查的方法,目前方法多集中于框架改进而忽略了数据预处理,文中所提方法旨在对球面纹理与结节形状进行向量表达,使其可以输入深度森林进行特征分类训练。方法 首先采用辽宁中医药大学附属医院数据,通过3维重构获得3维肺结节图像。其次使用球谐函数与容斥映射算法在保留空间信息的同时将纹理以网格方式映射到标准球面上。再次使用网格-LBP与映射形变能量分别完成对球面纹理与结节形状信息的构建。最后提出一种基于网格的多粒度扫描方法对深度森林训练框架进行改进,并将向量化后的纹理和形状特征加入到改进的深度森林训练框架中进行实验验证。结果 通过大量的实验结果验证,在准确率（ACC）、特异度（SPE）、敏感度（SEN）和受试者工作特征曲线下的面积（AUC）4个衡量指标下,本文方法具有优于现存先进方法的表现,其中ACC、SPE、SEN和AUC分别达到76.06%、69.46%、88.46%和0.84。结论 基于球谐函数与容斥映射算法可成功地对肺结节表面和形状两个特征进行向量化并训练,不仅考虑了数据预处理,而且通过两个特征对肺结节良恶性检测的准确率要高于传统1个特征检测的结果,同时也为3维模型中特征的提取及向量化提供了一个有效的方法。     

深度学习在肺结节辅助诊断中的应用

肺癌位居癌症死亡率首位,对其进行早期诊断和治疗可降低肺癌患者的死亡率。深度学习能够自动提取结节特征,并完成肺结节的良恶性及恶性等级分类,因此深度学习方法成为肺癌早期诊断的重要手段。对常用数据集进行介绍,系统阐述了栈式去噪自编码器（SDAE）、深度置信网络（DBN）、生成对抗网络（GAN）、卷积神经网络（CNN）、循环神经网络（RNN）和迁移学习技术在肺结节良恶性分类中的应用,阐述了深度卷积生成对抗网络（DCGAN）、多尺度卷积神经网络（MCNN）、U型网络（U-Net）和集成学习技术在肺结节恶性等级分类中的应用,针对肺结节分类的深度学习方法进行了综合分析,并对未来研究方向进行展望。     

基于注意力机制多任务的肺结节癌变风险判断方法

对于CT影像中检测出的肺部结节, 需要自动判断其是否有癌变风险. 不同于大多数现有的研究方法只区分结节良恶性, 本文提出了一个基于注意力机制的多任务学习模型, 将与结节良恶性相关的语义特征属性一并判断输出, 通过判断9个结节特征(对比度、分叶征、毛刺征、球形度、边缘、纹理、钙化程度、大小以及恶性程度)的同时实现内在特征的共享, 以达到提高各子任务性能的目的. 选择视觉转换器(ViT)模型作为多任务共享特征提取层, 整体模型采用动态加权平均方法来对各子任务的Loss函数进行优化. 在LUNA16数据集上的实验表明, 该学习框架可以提升肺结节癌变风险判断的性能, 且同时对其他语义特征的判断也能提升结果的可解释性.     

基于小样本学习的SAR图像识别

深度学习已成为图像识别领域的一个研究热点。与传统图像识别方法不同,深度学习从大量数据中自动学习特征,并且具有强大的自学习能力和高效的特征表达能力。但在小样本条件下,传统的深度学习方法如卷积神经网络难以学习到有效的特征,造成图像识别的准确率较低。因此,提出一种新的小样本条件下的图像识别算法用于解决SAR图像的分类识别。该算法以卷积神经网络为基础,结合自编码器,形成深度卷积自编码网络结构。首先对图像进行预处理,使用2D Gabor滤波增强图像,在此基础上对模型进行训练,最后构建图像分类模型。该算法设计的网络结构能自动学习并提取小样本图像中的有效特征,进而提高识别准确率。在MSTAR数据集的10类目标分类中,选择训练集数据中10%的样本作为新的训练数据,其余数据为验证数据,并且,测试数据在卷积神经网络中的识别准确率为76.38%,而在提出的卷积自编码结构中的识别准确率达到了88.09%。实验结果表明,提出的算法在小样本图像识别中比卷积神经网络模型更加有效。     

应用于CT图像肺结节检测的深度学习方法综述

肺癌是世界上死亡率最高的癌症，通过胸部CT影像检测肺结节对肺癌早期诊断和治疗意义重大。为了减轻放射科医生的工作量以及同时减少误诊率和漏诊率，研究人员提出了计算机辅助检测（CAD）系统辅助放射科医生检测和诊断肺结节。目前，研究人员正在尝试不同的深度学习技术，以提高计算机辅助诊断系统在基于CT图像的肺癌筛查中的性能。这项工作回顾了作为肺癌检测的CAD系统目前典型的深度学习的算法和框架，主要从数据集介绍、2D深度学习方法、3D深度学习方法、数据不平衡问题的处理、模型训练方法以及模型可解释性这六个方面进行介绍。最后，对各个方法的主要特点和算法性能进行了综合比较分析，并对如何提高结节检测性能进行了展望。     

基于主动深度学习的高光谱影像分类

针对当前高光谱遥感影像分类人工标注样本费时费力,大量未标注样本未得到有效利用以及主要利用光谱信息而忽视空间信息等问题,提出了一种空-谱信息与主动深度学习相结合的高光谱影像分类方法。首先利用主成分分析对原始影像进行降维,在此基础上提取像素的一正方形小邻域作为该像素的空间信息并结合其原始光谱信息得到空谱特征。然后,通过稀疏自编码器得到原始数据的稀疏特征表达,并通过逐层无监督学习稀疏自编码器构建深度神经网络,输出原始数据的深度特征,将其连接到softmax分类器,利用少量标记样本以监督学习的方式完成模型的精调。最后,利用主动学习算法选择最不确定性样本对其进行标注,并加入至训练样本以提高分类器的分类效果。分别对PaviaU影像和PaviaC影像进行分类实验的结果表明,该方法在少量标记样本情况下,相对于传统方法能有效地提高分类精度。     

Latent Space Representational Learning of Deep Features for Acute Lymphoblastic Leukemia Diagnosis

Acute Lymphoblastic Leukemia (ALL) is a fatal malignancy that is featured by the abnormal increase of immature lymphocytes in blood or bone marrow. Early prognosis of ALL is indispensable for the effectual remediation of this disease. Initial screening of ALL is conducted through manual examination of stained blood smear microscopic images, a process which is time-consuming and prone to errors. Therefore, many deep learning-based computer-aided diagnosis (CAD) systems have been established to automatically diagnose ALL. This paper proposes a novel hybrid deep learning system for ALL diagnosis in blood smear images. The introduced system integrates the proficiency of autoencoder networks in feature representational learning in latent space with the superior feature extraction capability of standard pretrained convolutional neural networks (CNNs) to identify the existence of ALL in blood smears. An augmented set of deep image features are formed from the features extracted by GoogleNet and Inception-v3 CNNs from a hybrid dataset of microscopic blood smear images. A sparse autoencoder network is designed to create an abstract set of significant latent features from the enlarged image feature set. The latent features are used to perform image classification using Support Vector Machine (SVM) classifier. The obtained results show that the latent features improve the classification performance of the proposed ALL diagnosis system over the original image features. Moreover, the classification performance of the system with various sizes of the latent feature set is evaluated. The retrieved results reveal that the introduced ALL diagnosis system superiorly compete the state of the art.     

基于深度学习模型的无线传感器网络数据融合算法

为了在无线传感器网络中提高数据融合性能,基于深度学习模型,提出一种将层叠自动编码器(SAE)和分簇协议相结合的数据融合算法SAEMDA,该算法在各个簇内构建特征提取分类模型SAEM,通过SAEM对节点数据进行特征提取和分类,之后将同类特征融合并发送给汇聚节点。SAEM的训练既可以采用离线有监督学习也可以采用在线无监督学习。仿真实验表明:和BPFDA,SOFMDA算法相比,SAEMDA在网络能耗大致相当的情况下能将数据融合正确率提高最多7.5%。     

多标记学习自编码网络无监督维数约简

多标记学习是针对一个实例同时与一组标签相关联而提出的一种机器学习框架,是该领域研究热点之一,降维是多标记学习一个重要且具有挑战性的工作。针对有监督的多标记维数约简方法,提出一种无监督自编码网络的多标记降维方法。首先,通过构建自编码神经网络,对输入数据进行编码和解码输出;然后,引入稀疏约束计算总体成本,使用梯度下降法进行迭代求解;最后,通过深度学习训练获得自编码网络学习模型,提取数据特征实现维数约简。实验中使用多标记算法ML-kNN做分类器,在6个公开数据集上与其他4种方法对比。实验结果表明,该方法能够在不使用标记的情况下有效提取特征,降低多标记数据维度,稳定提高多标记学习性能。     

Improving the performance of lung nodule classification by fusing structured and unstructured data

We investigated the issue of improving the classification performance for pulmonary nodules by learning the fusion features of structured and unstructured data. Current strategies for lung nodule classification, such as radiomics methods and deep learning approaches, all share the flaw of only using the unstructured data of patients, which is always a collection of medical images (e.g., computed tomography (CT) scans, X-rays, and pathological sections), while ignoring the structured data (e.g., baseline demographics, clinical characteristics, and laboratory examinations). However, from a clinical perspective, all of this information is required for accurate patient diagnosis. Therefore, to exploit all patient information, we addressed a more difficult problem: jointly modeling the multimodal patient data. Two models are proposed to combine structured and unstructured data. One employs deep learning with a softmax classifier (the structured and unstructured data fusion neural network (SUDFNN)), and the other implements an extreme gradient boosting (XGBoost) classifier (the structured and unstructured data fusion XGBoost (SUDFX)). The annotated structured data in the extensible markup language (XML) file from the Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) database and the CT scans from the LUng Nodule Analysis 2016 (LUNA16) dataset were used to validate our model. The results show that the performance of the model is significantly improved when introducing the structured data, regardless of the nodule cube size and which classifier is used. The rationale for the improvement with the addition of structured features is provided. The optimal accuracy, sensitivity, specificity, and area under the receiver operating characteristic (ROC) curve (AUC) values reached 0.936, 0.919, 0.956, and 0.971, respectively. Consequently, fusing structured and unstructured data can uncover more patient information and provide better decision support for the clinical diagnosis and treatment process, providing good application value and promotion prospects.     

Discriminative Feature Learning via Sparse Autoencoders with Label Consistency Constraints

Autoencoders have been successfully used to build deep hierarchical models of data. However, a deep architecture usually needs further supervised fine-tuning to obtain better discriminative capacity. To improve the discriminative capacity of deep hierarchical features, this paper proposes a new deterministic autoencoder, trained by a label consistency constraints algorithm that injects discriminative information to the network. We introduce the center loss as label consistency constraints to learn the hidden features of data and add it to the Sparse AutoEncoder to form a new autoencoder, namely Label Consistency Constrained Sparse AutoEncoders (LCCSAE). Specifically, the center loss learns the center of each class, and simultaneously penalizes the distances between the features and their corresponding class centers. In the end, autoencoders are stacked to form a deep architecture of LCCSAE for image classification tasks. To validate the effectiveness of LCCSAE, we compare it with other autoencoders in terms of the deeply learned features and the subsequent classification tasks on MNIST and CIFAR-bw datasets. Experimental results demonstrate the superiority of LCCSAE over other methods.     

Classification and recognition of computed tomography images using image reconstruction and information fusion methods

In this paper, we propose a diagnosis and classification method of hydrocephalus computed tomography (CT) images using deep learning and image reconstruction methods. The proposed method constructs pathological features differing from the other healthy tissues. This method tries to improve the accuracy of pathological images identification and diagnosis. Identification of pathological features from CT images is an essential subject for the diagnosis and treatment of diseases. However, it is difficult to accurately distinguish pathological features owing to the variability of appearances, fuzzy boundaries, heterogeneous densities, shapes and sizes of lesions, etc. Some study results reported that the ResNet network has a better classification and diagnosis performance than other methods, and it has broad application prospectives in the identification of CT images. We use an improved ResNet network as a classification model with our proposed image reconstruction and information fusion methods. First, we evaluate a classification experiment using the hydrocephalus CT image datasets. Through the comparative experiments, we found that gradient features play an important role in the classification of hydrocephalus CT images. The classification effect of CT images with small information entropy is excellent in the evaluation of hydrocephalus CT images. A reconstructed image containing two channels of gradient features and one channel of LBP features is very effective in classification. Second, we apply our proposed method in classification experiments on CT images of colonography polyps for an evaluation. The experimental results have consistency with the hydrocephalus classification evaluation. It shows that the method is universal and suitable for classification of CT images in these two applications for the diagnosis of diseases. The original features of CT images are not ideal characteristics in classification, and the reconstructed image and information fusion methods have a great effect on CT images classification for pathological diagnosis.