基于多模态MRI与深度学习的乳腺病变良恶性鉴别

为提高基于动态增强磁共振成像（DCE-MRI）的计算机辅助（CAD）方法对乳腺病变良恶性鉴别的精度,本文基于多模态特征融合,提出一种联合非对称卷积和超轻子空间注意模块的卷积神经网络AC_Ulsam_CNN.首先,采用迁移学习方法预训练模型,筛选出对乳腺病变良恶性鉴别最为有效的DCE-MRI扫描时序.而后,基于最优扫描时序图像,搭建基于AC_Ulsam_CNN网络的模型,以增强分类模型的特征表达能力和鲁棒性.最后,将影像特征与乳腺影像数据报告和数据系统（BI-RADS）分级、表观扩散系数（ADC）和时间-信号强度曲线（TIC）类型等多模态信息进行特征融合,以进一步提高模型对病灶的预测性能.采用五折交叉验证方法进行模型验证,本文方法获得了0.826的准确率（ACC）和0.877的受试者工作曲线下面积（AUC）.这表明该算法在小样本量数据下可较好区分乳腺病变的良恶性,而基于多模态数据的融合模型也进一步丰富了特征信息,从而提高病灶的检出精度,为乳腺病灶良恶性的自动鉴别诊断提供了新方法.     

基于脑部磁共振图像三维局部模式变换特征提取进行阿尔茨海默病病程预测分类

本文提出一种三维局部模式变换提取进行纹理特征并与常规特征相融合的方法,基于脑部磁共振图像,对认知功能正常的健康人体（CN）、轻度认知障碍（MCI）患者和阿尔茨海默病（AD）患者进行预测分类.首先对46例CN对照组、61例MCI患者和25例AD患者的脑部磁共振图像提取感兴趣区域,然后提取双侧海马体组织、灰质和白质的三维局部模式变换纹理特征和常规特征,并将两类特征融合,使用支持向量机分类算法进行分类.结果显示利用本方法,基于双侧海马体组织对AD组和CN组进行分类的准确率为88.73%、敏感度为78.00%、特异度为95.7%、受试者工作特征（ROC）曲线下面积（AUC）为0.886 5;基于灰质的准确率为85.92%、敏感度为80.00%、特异度为86.6%、AUC为0.854 3.这证明基于海马体磁共振图像,利用本文提出的改进三维局部模式变换提取的纹理特征进行阿尔茨海默病病程分类效果较好,融合常规特征后更可提高分类预测的精度.     

基于深度学习的胰腺黏液性和浆液性囊性肿瘤的多源特征分类模型

术前精准预测胰腺囊性肿瘤的类型,对制定个体化诊疗方案具有重要的临床价值．针对胰腺浆液性和黏液性囊性肿瘤的分类鉴别问题,本文探讨了基于深度学习的多源特征分类模型在胰腺囊性肿瘤的术前辅助诊断中的应用．首先,通过深度学习和影像组学技术从分割图像中提取深度学习特征和影像组学特征,并对病例的临床特征进行评估和量化,然后采用最小绝对收缩选择算子（LASSO）及交叉验证的方法筛选特征,随之构建出两个多源特征模型,即影像组学联合深度学习（RAD＿DL）模型、临床特征联合RAD＿DL(Clinical＿RAD＿DL)模型,把传统的影像组学（RAD）模型和深度学习（DL）模型作为对照,最后选用支持向量机（SVM）、自适应提升算法（ADAboost）、随机森林（Random Forest）以及逻辑回归（Logistic）进行分类．采用准确率、召回率、精确率、曲线下面积（AUC）值以及精确率和召回率的调和平均数（F1值）作为评价指标,比较上述4种不同特征模型的分类效能,用校准曲线和决策曲线来评估其临床应用价值．结果显示Clinical＿RAD＿DL特征模型的分类效能表现最佳,准确率是0.923 1,召回率是0...     

基于磁共振图像和改进的UNet++模型区分阿尔茨海默症患者和健康人群

阿尔茨海默症（Alzheimer's Disease，AD）是一种神经退行性疾病，高效准确的早期诊断对其治疗至关重要．本文提出了一种融合多语义级别的深度卷积神经网络结构，基于磁共振图像，用于区分AD患者与正常受试者的方法．首先，在传统UNet++网络的基础上改进了深度监督整合算法；然后，构建了一种新的特征融合结构，进一步细化了不同语义级别的特征；最后，基于不同组织区域（白质、灰质和脑脊液）的磁共振图像，使用本文所提出的方法区分AD患者和正常受试者，并探究了从不同组织获得的信息对分类准确率的影响．实验结果表明，使用本文提出的方法区分两类人群的最高准确率为98.74%，平均准确率为98.47%，高于目前文献报道的其他方法．     

非线性超声射频信号熵对乳腺结节良恶性的定征

本文提出了一种基于非线性超声射频(radio frequency, RF)信号熵对乳腺结节良恶性进行定征的方法.对306例乳腺结节样本(良性158例,恶性148例)提取了基于超声RF信号二次谐波的熵和加权熵,以及常规超声参数(图像灰度、纵横比、不规则度、乳腺结节大小、深度);采用t检验和线性分类器检测参数对乳腺结节良恶性的区分度;进一步将有效参数组合输入支持向量机对乳腺结节良恶性进行分类.结果表明:除图像灰度外,其余参数均在乳腺结节的良性与恶性间有显著差异.多参数结合输入支持向量机的良恶性分类的准确率、敏感性和特异性分别为81.4%, 78.4%和84.2%.本文工作表明非线性超声RF信号的熵可有效地定征乳腺结节的良恶性,有望成为乳腺结节良恶性定征新参量.     

基于新型支持向量机的影像组学在肝脏结节分类中的应用

肝癌是最常见的恶性肿瘤之一,亚洲地区最为常见的肝癌演变过程为肝炎-肝硬化结节-异型增生结节-肝细胞性肝癌．判断肝脏结节在演变过程所处分期,并采取干预措施,对降低肝癌的发生率非常关键．本文针对影像组学提出了更精确的支持向量机（SVM）分类算法——LFOA-F-SVM,用于对120名患者的腹部动态增强磁共振图像的肝脏结节进行四分类．该算法利用了考虑半径与几何间距的F-SVM,并结合莱维飞行策略（LF）的果蝇优化算法（FOA）寻求超参．为了验证方法的有效性,本文另外添加了5个UCI分类数据集（心脏、帕金森疾病、虹膜、葡萄酒和动物园）,并与SVM、PSO-SVM、FOA-SVM、F-SVM进行比较．结果表明,在6个分类数据集（包括肝脏结节数据集和5个UCI分类数据集）中,相对于其他分类算法,LFOA-F-SVM的分类准确率最高,在肝脏结节数据集中的四分类精确率和查全率也较高．     

基于影像组学的直肠癌术前T分期预测

直肠癌T分期对患者的术前评估有重要作用.然而,传统的放射科医生根据患者磁共振图像直接判断分期的方法效果欠佳.本文提出使用影像组学的方法预测直肠癌T分期,首先获取105例直肠癌患者影像数据,根据病理报告中的T分期结果将T1、T2期患者划分为未突破肌层组,将T3、T4期患者分为突破肌层组,整理数据得到未突破肌层组31例,突破肌层组74例.在患者的轴向位T2WI图像中勾画病灶区域,并在病灶上使用pyradiomics工具包提取影像组学特征,使用最小绝对值收敛和选择算子（LASSO）对高维特征做特征选择,得到与T分期高度相关的特征数据,使用随机森林、支持向量机（SVM）、逻辑回归、梯度提升树（GBDT）分别建模,进行交叉验证调参,评估模型性能.每层图像提取100维特征,经LASSO特征选择后得到7个与T分期高度相关的特征,使用4种模型分别建模,其中SVM算法表现最优,平均受试者操作特征曲线下面积（AUC）、准确率、灵敏度、特异度分别为0.968 5、0.886 4、0.962 5、0.899 2,测试集准确率达到了0.904 7.结果表明,使用影像组学方法可以提高直肠癌T分期预测的准确率.     

结合小波融合和深度学习的脑胶质瘤自动分割

针对水肿区域边界模糊和瘤内结构复杂多变导致的脑胶质瘤分割不精确问题，本文提出了一种基于小波融合和3D-UNet网络的脑胶质瘤磁共振图像自动分割算法.首先，对脑胶质瘤磁共振图像的T1、T1ce、T2、Flair四种模态进行小波融合以及偏置场校正；然后，提取待分类的图像块；再利用提取的图像块训练3D-UNet网络以对图像块中的像素进行分类；最后加载损失率较小的网络模型进行分割，并采用基于连通区域的轮廓提取方法，以降低假阳性率.对57组Brats2018（Brain Tumor Segmentation 2018）磁共振图像测试集进行分割的结果显示，肿瘤的整体、核心和水肿部分的平均分割准确率（DSC）分别达到90.64%、80.74%和86.37%，这表明该算法分割脑胶质瘤准确率较高，与金标准相近.相比多模态图像融合前，该算法在减少输入网络数据量和图像冗余信息的同时，还一定程度上解决了胶质瘤边界模糊、分割不精确的问题，提高了分割的准确度和鲁棒性.     

基于OPO脉冲激光激发光声光谱的真假血液分类鉴别

为了实现快速准确并可回收再利用地鉴别真血和假血,采用光声光谱技术构建了一套血液光声检测系统并获取血样的光声信号。选取三种动物真血（马血、牛血和兔血）和两种假血（道具假血和红墨汁）共125组血样作为实验样本。获取了700～1 064 nm波段内所有样本的光声信号和光声峰峰值谱。实验表明,真血和假血的光声信号幅度、轮廓、峰值时间点和光声峰峰值均存在差异。为了实现高准确度的真假血液分类识别,采用了遗传优化的小波神经网络（WNN-GA）算法,对全波段100组样本进行训练,并构建了类Morlet小波基函数,然后对25组测试血样进行分类识别。利用遗传算法对WNN网络的权值、阈值和小波基函数平移、伸缩因子进行了优化,同时通过调节两个学习率因子,将真假血液的分类识别率提高了24%。采用主成分分析（PCA）对全波段血样光声峰峰值进行特征提取,再利用WNN-GA算法进行训练和分类识别。结果表明,在主成分个数为6时,PCA-WNN-GA融合算法可以使真假血液的分类识别率提高到100%。与另外6种分类识别算法相对比,该融合算法的识别准确率明显占优。光声光谱技术联合PCA-WNN-GA算法,可以准确地实现真假血液的分类鉴别。     

主成分分析排序和模糊线性判别分析的生菜近红外光谱分类

贮存时间是影响生菜品质的一项重要因素,传统的贮存时间鉴别方法主要依靠人工经验,但是这种方法的准确率和可信度并不高。研究的目标是建立一种基于模糊识别的模型进行生菜光谱分析以实现生菜贮存时间的鉴别,并与其他鉴别方法作比较。为此,在当地超市购买60份新鲜生菜样品,存放于冰箱中待用。首先,通过Antaris Ⅱ近红外光谱检测仪采集生菜样品的近红外光谱数据,每隔12小时检测一次,每个样本检测重复三次,并取三次平均值作为实验数据。其次,利用多元散射校正（MSC）减少近红外光谱中的冗余信息。为了进一步去除近红外光谱中的无用信息以及简化随后的数据分类过程,分别运用主成分分析（PCA）和排序主成分分析 (PCA Sort)。其中,PCA Sort通过改进对主成分的排序方法能提高分类准确率,同时便于模糊线性鉴别分析（FLDA）进一步提取特征。PCA和PCA Sort的计算仅运用了前15个主成分（能充分反映光谱的主要信息）。最后,利用模糊线性鉴别分析算法（FLDA）和K近邻算法（KNN）进一步分类所得的低维数据。基于PCA和KNN算法的模型鉴别准确率达到43%,而基于PCA,FLDA和KNN算法的模型鉴别准确率可达83%。上述结果说明基于PCA,FLDA和KNN算法的模型鉴别准确率已经得到较大程度提高。当用PCA Sort替代了模型中的PCA算法后,结合FLDA和KNN算法则鉴别准确率达到98.33%。实验结果表明PCA Sort结合FLDA和KNN所建立的模型是有效的生菜贮存时间鉴别模型。     

A knowledge-driven feature learning and integration method for breast cancer diagnosis on multi-sequence MRI

BackgroundThe classification of benign versus malignant breast lesions on multi-sequence Magnetic Resonance Imaging (MRI) is a challenging task since breast lesions are heterogeneous and complex. Recently, deep learning methods have been used for breast lesion diagnosis with raw image input. However, without the guidance of domain knowledge, these data-driven methods cannot ensure that the features extracted from images are comprehensive for breast cancer diagnosis. Specifically, these features are difficult to relate to clinically relevant phenomena.PurposeInspired by the cognition process of radiologists, we propose a Knowledge-driven Feature Learning and Integration (KFLI) framework, to discriminate between benign and malignant breast lesions using Multi-sequences MRI.MethodsStarting from sequence division based on characteristics, we use domain knowledge to guide the feature learning process so that the feature vectors of sub-sequence are constrained to lie in characteristic-related semantic space. Then, different deep networks are designed to extract various sub-sequence features. Furthermore, a weighting module is employed for the integration of the features extracted from different sub-sequence images adaptively.ResultsThe KFLI is a domain knowledge and deep network ensemble, which can extract sufficient and effective features from each sub-sequence for a comprehensive diagnosis of breast cancer. Experiments on 100 MRI studies have demonstrated that the KFLI achieves sensitivity, specificity, and accuracy of 84.6%, 85.7% and 85.0%, respectively, which outperforms other state-of-the-art algorithms.     

Handcrafted MRI radiomics and machine learning: Classification of indeterminate solid adrenal lesions

PurposeTo assess a radiomic machine learning (ML) model in classifying solid adrenal lesions (ALs) without fat signal drop on chemical shift (CS) as benign or malignant.Method55 indeterminate ALs (21 lipid poor adenomas, 15 benign pheocromocytomas, 1 oncocytoma, 12 metastases, 6 primary tumors) showing no fat signal drop on CS were retrospectively included. Manual 3D segmentation on T2-weighted and CS images was performed for subsequent radiomic feature extraction. After feature stability testing and an 80–20% train-test split, the train set was balanced via oversampling. Following a multi-step feature selection, an Extra Trees model was tuned with 5-fold stratified cross-validation in the train set and then tested on the hold-out test set.ResultsA total of 3396 features were extracted from each AL, of which 133 resulted unstable while none had low variance (< 0.01). Highly correlated (r > 0.8) features were also excluded, leaving 440 parameters. Among these, Support Vector Machine 5-fold stratified cross-validated recursive feature elimination selected a subset of 6 features. ML obtained a cross-validation accuracy of 0.94 on the train and 0.91 on the test sets. Precision, recall and F1 score were respectively 0.92, 0.91 and 0.91.ConclusionsOur MRI handcrafted radiomics and ML pipeline proved useful to characterize benign and malignant solid indeterminate adrenal lesions.     

Relationship between lesion size and signal enhancement on subtraction fat-suppressed MR imaging of the breast

OBJECTIVE: To investigate the relationship between size and whole lesion enhancement of breast neoplasms. MATERIALS AND METHODS: Fat-suppressed subtraction MRI was performed in 94 breast lesions (44 malignant, 50 benign) with pathologically confirmed diagnoses. Of these, all malignant lesions and 31 of the 50 benign lesions showed enhancement. The degree of enhancement was quantified by using an ROI tracing around the whole lesion and calculated as the percentage increase in signal intensity between the corresponding precontrast and postcontrast images. RESULTS: The 44 malignant lesions showed enhancement percentage of 38.3% to 186.4% (mean 109.9%), and the 31 benign lesions showed enhancement percentage of 12.8% to 180.2% (mean 79.5%). The difference is statistically significant (P = .002). In 54 small lesions (28 malignant, 26 benign) with enhancing pixel areas of <300 mm(2) corresponding to a diameter of approximately 19.5 mm, an enhancement exceeding 75% of baseline separated malignant lesions (mean enhancement 116.7%) from benign ones (mean enhancement 72.8%) (P = .0001). This gave a sensitivity of 100% and a specificity of 69%, a positive predictive value of 78%, negative predictive value of 100% and an accuracy of 85% in using >75% enhancement increase in detecting malignancy in small (<300 mm(2)) enhancing lesions. CONCLUSION: The high sensitivity in the detection of small malignant lesions suggests a potential for the method to be used in assessment of small enhancing breast lesions.     

Diagnosis of suspicious breast lesions using an empirical mathematical model for dynamic contrast-enhanced MRI

The purpose of this study was to test whether an empirical mathematical model (EMM) of dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) can distinguish between benign and malignant breast lesions. A modified clinical protocol was used to improve the sampling of contrast medium uptake and washout. T(1)-weighted DCE magnetic resonance images were acquired at 1.5 T for 22 patients before and after injection of Gd-DTPA. Contrast medium concentration as a function of time was calculated over a small region of interest containing the most rapidly enhancing pixels. Then the curves were fitted with the EMM, which accurately described contrast agent uptake and washout. Results demonstrate that benign lesions had uptake (P<2.0 x 10(-5)) and washout (P<.01) rates of contrast agent significantly slower than those of malignant lesions. In addition, secondary diagnostic parameters, such as time to peak of enhancement, enhancement slope at the peak and curvature at the peak of enhancement, were derived mathematically from the EMM and expressed in terms of primary parameters. These diagnostic parameters also effectively differentiated benign from malignant lesions (P<.03). Conventional analysis of contrast medium dynamics, using a subjective classification of contrast medium kinetics in lesions as "washout," "plateau" or "persistent" (sensitivity=83%, specificity=50% and diagnostic accuracy=72%), was less effective than the EMM (sensitivity=100%, specificity=83% and diagnostic accuracy=94%) for the separation of benign and malignant lesions. In summary, the present research suggests that the EMM is a promising alternative method for evaluating DCE-MRI data with improved diagnostic accuracy.     

Deep-learning approach with convolutional neural network for classification of maximum intensity projections of dynamic contrast-enhanced breast magnetic resonance imaging

PurposeWe aimed to evaluate deep learning approach with convolutional neural networks (CNNs) to discriminate between benign and malignant lesions on maximum intensity projections of dynamic contrast-enhanced breast magnetic resonance imaging (MRI).MethodsWe retrospectively gathered maximum intensity projections of dynamic contrast-enhanced breast MRI of 106 benign (including 22 normal) and 180 malignant cases for training and validation data. CNN models were constructed to calculate the probability of malignancy using CNN architectures (DenseNet121, DenseNet169, InceptionResNetV2, InceptionV3, NasNetMobile, and Xception) with 500 epochs and analyzed that of 25 benign (including 12 normal) and 47 malignant cases for test data. Two human readers also interpreted these test data and scored the probability of malignancy for each case using Breast Imaging Reporting and Data System. Sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve (AUC) were calculated.ResultsThe CNN models showed a mean AUC of 0.830 (range, 0.750–0.895). The best model was InceptionResNetV2. This model, Reader 1, and Reader 2 had sensitivities of 74.5%, 72.3%, and 78.7%; specificities of 96.0%, 88.0%, and 80.0%; and AUCs of 0.895, 0.823, and 0.849, respectively. No significant difference arose between the CNN models and human readers (p > 0.125).ConclusionOur CNN models showed comparable diagnostic performance in differentiating between benign and malignant lesions to human readers on maximum intensity projection of dynamic contrast-enhanced breast MRI.     

Diagnostic utility of tumor vascularity on magnetic resonance imaging of the breast

The objective of this study was to examine the relation of tumor vascularity on magnetic resonance imaging (MRI) with differential diagnosis malignant from benign lesions and tumor invasiveness in breast carcinoma. Forty-nine patients with breast cancer or benign lesion (median 49 yrs) were examined with dynamic MRI. Scanning of the entire breast was performed at 1.5 T with a three-dimensional fast spin echo sequence, using an original polarity altered spectral and spatial selective acquisition (PASTA) technique for fat suppression. Subsequently 0.1 mmol/Kg Gd-DTPA was administered and 3-6 images were obtained. The presence or absence of intratumoral, marginal and peritumoral vascularity on MRI was recorded. The excised specimen was histopathologically examined for the size of lesion, the presence and extent of local invasion. Tumor size on MRI correlated closely with the size at morphologic examination (r = 0. 96). Intratumoral (p = 0.04), marginal (p = 0.05) and peritumoral vascularity (p = 0.01) were less common in benign than in malignant lesions. Among the latter, intratumoral (p = 0.01) and marginal (p = 0.03) vascularity were more common in invasive carcinomas than in DCIS. In the subset of invasive carcinomas (n = 31); however, the tumors exhibiting intratumoral vascularity were markedly larger (p = 0.03). The presence of intratumoral and marginal vascularity on MRI can help predict both the differential diagnosis malignant from benign lesions and the presence tumor invasion in breast carcinomas.