贝叶斯概率图像自动分割研究

探讨了一种新的图像自动分割的方法。提出应用高斯有限混合模型与期望－极大化算法对图像特征空间的数据进行聚类，采用信息理论准则（ITC）确定要分割的图像区域数目，用贝叶斯概率分割图像。整合这些技术可以实现图像自动分割，而且实验结果表明信息理论准则可以确定适当的区域数目。     

基于互信息改进型脉冲耦合神经网络多值图像分割

为了自动地进行图像的多值分割,从原始图像与分割图像之间的相互关系出发,以最大互信息为优化分割目标,以互信息熵差作为一种新的分类类数判据,在对传统脉冲耦合神经网络模型改进的基础上,提出了一种基于最大互信息改进型脉冲耦合神经网络图像多值分割算法.理论分析和实验结果表明,该方法能够自动确定最佳分割迭代次数及最佳分割灰度类数,对分割图像具有良好的特征划分能力,且在分割类数较少的情况下,能较好地保持图像细节、纹理及边缘等信息,对不同图像分割准确度高,具有较强的适用性.     

基于互信息改进型脉冲耦合神经网络多值图像分割

为了自动地进行图像的多值分割,从原始图像与分割图像之间的相互关系出发,以最大互信息为优化分割目标,以互信息熵差作为一种新的分类类数判据,在对传统脉冲耦合神经网络模型改进的基础上,提出了一种基于最大互信息改进型脉冲耦合神经网络图像多值分割算法.理论分析和实验结果表明,该方法能够自动确定最佳分割迭代次数及最佳分割灰度类数,对分割图像具有良好的特征划分能力,且在分割类数较少的情况下,能较好地保持图像细节、纹理及边缘等信息,对不同图像分割准确度高,具有较强的适用性.     

基于3D卷积神经网络的脑肿瘤图像分割

三维脑胶质瘤磁共振成像肿瘤形状各异、边缘模糊,目前大多数基于2D卷积神经网络的分割方法不能很好的分割三维图像。为了能够准确分割出三维图像中的肿瘤部分,提出一种融合多尺度特征信息的3D卷积神经网络脑肿瘤图像分割方法。利用并行的3D空洞卷积提取特征信息,将不同感受野的信息融合。将Dice损失和BCE损失结合,形成一种新的损失函数并配合恒等映射,进一步提高分割精度。在BraTs2020数据集上对模型进行验证,结果表明,该模型分割的全肿瘤区、核心区和增强区的Dice系数分别为89.1%、83.9%和82.6%。在LGG脑部肿瘤图像数据集上对模型进行验证,结果表明,Dice系数达到了93.3%。所提出的分割方法不仅能够精确的分割三维脑胶质瘤图像,而且同样适用于分割二维脑胶质瘤图像。     

用于SAR估计的基于U-Net网络的快速膝关节模型重建

膝关节高场磁共振成像（MRI）时,射频功率沉积（SAR）是一个关键的安全指标.目前对于局部SAR的准确估计只能通过电磁仿真实现,这就要求得到每一个个体的膝关节模型.本文提出一种针对低场磁共振图像的基于卷积神经网络的分割方法,以实现膝关节磁共振图像的快速重建.数据集来自于矢位T₁加权自旋回波图像,将膝关节组织按照"肌肉-脂肪-骨骼"模型进行简化,除脂肪与骨骼之外的其他组织归类为肌肉.采用一种全卷积的神经网络,即U-Net进行逐层的图像分割,卷积层数为4,训练采用交叉熵函数.本文对图像的自动分割结果与手动标注结果进行了定量的比较.此外,采用3 T正交鸟笼线圈进行了SAR仿真,结果验证了组织简化对于SAR估计的可行性,并且所提方法构建的模型可以得到较为精准的局部SAR分布.     

传统解剖学特征与深度学习相结合的肺叶分割算法

CT图像中肺叶位置的确定对于肺部疾病的准确定位以及定性定量分析具有重要意义。为了提高肺叶自动分割准确率,提出了一种结合气管,血管等传统解剖学特征以及深度学习的肺叶分割算法。对原始图像进行预处理,获取肺实质、气管、血管以及基于深度学习网络的肺裂分割结果;整合来自多个解剖结构的信息生成分水岭分割所需成本图像;通过基于深度学习网络的肺叶粗分割结果,获取肺叶标记区域;执行基于标记的分水岭分割,实现肺叶的自动分割。选取了来自上海市肺科医院的20例含有肺部疾病患者的CT图像对该方法进行验证,最终的Jaccard相似性系数为92.4%。实验结果表明方法具有较高的肺叶分割精度,并且具有较强的鲁棒性。     

基于水平集的闪光照相图像分割算法

针对Chan-Vese(CV)模型局部控制能力差的缺点,将基于区域的CV模型和分割曲线的局部信息结合起来,提出了一种新的水平集图像分割算法。该算法以CV法的分割曲线为初始曲线,以获得全局收敛性,在后继分割中引入分割曲线的局部信息,以提高模型对图像中微弱信号的分割能力。对闪光照相图像的数值实验表明,该算法噪声抵抗能力强,对初始轮廓位置不敏感,能实现对含细长拓扑结构和微小孔洞的弱边界闪光图像的自动分割。     

面向机器人手术模拟器的神经辐射场软组织动态三维重建

构建了一种基于自监督的框架,该框架从单目立体内窥镜视频中提取多视图图像,利用图像中的底层三维（3D）信息构建对象的几何约束,实现软组织结构的准确重建。基于分割任意场景模型对内窥镜下的动态手术器械、静态腹腔场景及可形变软组织结构进行分割解耦。该框架利用简单的神经网络多层感知机来表示动态神经辐射场（NeRF）中运动手术器械和形变软组织结构,基于偏斜熵损失对手术场景中的手术器械、腔体场景和软组织结构进行正确分离。在通过使用单目立体内窥镜捕获机器人手术模拟器场景的数据集上,将所提方法的结果与其他方法进行定量定性比较。结果表明本文方法在处理腹腔体场景、软组织结构重建、手术器械的分割解耦,以及来自多视点的3D信息和运动对象的图像分割等方面显著优于当前的方法。     

基于水平集的闪光照相图像分割算法

针对Chan-Vese(C-V)模型局部控制能力差的缺点,将基于区域的C-V模型和分割曲线的局部信息结合起来,提出了一种新的水平集图像分割算法。该算法以C-V法的分割曲线为初始曲线,以获得全局收敛性,在后继分割中引入分割曲线的局部信息,以提高模型对图像中微弱信号的分割能力。对闪光照相图像的数值实验表明,该算法噪声抵抗能力强,对初始轮廓位置不敏感,能实现对含细长拓扑结构和微小孔洞的弱边界闪光图像的自动分割。     

基于SPCNN与改进型矢量CV模型的乳腺X射线肿块分割方法

作为乳腺癌计算机辅助诊断系统的重要环节,肿块分割的结果严重影响到肿块良恶性的判别.针对现有方法的不足,本文提出了一种基于简化型脉冲耦合神经网络和改进型矢量无边缘活动轮廓模型的乳腺X射线肿块分割方法.首先,通过数学分析计算SPCNN的相关参数与终止条件,进而利用SPCNN模型分割出肿块的初始轮廓.然后,针对传统CV模型的不足,进行相应的修正得到改进型矢量CV模型.最后,结合SPCNN分割出的初始轮廓,利用改进型的矢量CV模型处理ROI分割出肿块.采用北京大学人民医院乳腺中心提供的临床图像以及DDSM数据库的图像进行对比实验,实验结果表明,本文方法相比较现有方法分割结果更为准确,尤其是在处理东方女性致密性案例时,本文方法更有优势.     

Segmentation of whole breast and fibroglandular tissue using nnU-Net in dynamic contrast enhanced MR images

PurposeSegmentation of the whole breast and fibroglandular tissue (FGT) is important for quantitatively analyzing the breast cancer risk in the dynamic contrast-enhanced magnetic resonance (DCE-MR) images. The purpose of this study is to improve the accuracy and efficiency of the segmentation of the whole breast and FGT in 3-D fat-suppressed DCE-MR images with a versatile deep learning (DL) framework.MethodsWe randomly collected 100 breast DCE-MR scans from Shanghai Cancer Hospital of Fudan University. The MR scans in the dataset were different in both the spatial resolution and the MR scanners employed. Furthermore, four breast density categories were assessed by radiologists based on Breast Imaging Reporting and Data System (BI-RADS) of American College of Radiology. The dataset was separated into the training and the testing sets, while keeping a balanced distribution of scans with different imaging parameters and density categories. The nnU-Net has been recently proposed to automatically adapt preprocessing strategies and network architectures for a given medical image dataset, thus showing a great potential in the systematic adaptation of DL methods to different datasets. In this study, we applied the nnU-Net to segment the whole breast and FGT in 3-D fat-suppressed DCE-MR images. Five-fold cross validation was employed to train and validate the segmentation method.ResultsThe segmentation performance was evaluated with the volume and surface agreement metrics between the DL-based automatic and the manually delineated masks, as quantified with the following measures: the average Dice volume overlap (0.968 ± 0.017 and 0.877 ± 0.081), the average surface distances (0.201 ± 0.080 mm and 0.310 ± 0.043 mm), and the Pearson correlation coefficient of masks (0.995 and 0.972) between the automatic and the manually delineated masks, as calculated for the whole breast and the FGT segmentation, respectively. The correlation coefficient between the breast densities obtained with the DL-based segmentation and the manual delineation was 0.981. There was a positive bias of 0.8% (DL-based relative to manual) in breast density measurement with the Bland-Altman plot. The execution time of the DL-based segmentation was approximately 20 s for the whole breast segmentation and 15 s for the FGT segmentation.ConclusionsOur DL-based segmentation framework using nnU-Net could robustly achieve high accuracy and efficiency across variable MR imaging settings without extra pre- or post-processing procedures. It would be useful for developing DCE-MR-based CAD systems to quantify breast cancer risk and to be integrated into the clinical workflow.     

Computer-aided diagnosis of breast DCE-MRI using pharmacokinetic model and 3-D morphology analysis

Three-dimensional (3-D) dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) consists of a large number of images in different enhancement phases which are used to identify and characterize breast lesions. The purpose of this study was to develop a computer-assisted algorithm for tumor segmentation and characterization using both kinetic information and morphological features of 3-D breast DCE-MRI. An integrated color map created by intersecting kinetic and area under the curve (AUC) color maps was used to detect potential breast lesions, followed by the application of a region growing algorithm to segment the tumor. Modified fuzzy c-means clustering was used to identify the most representative kinetic curve of the whole segmented tumor, which was then characterized by using conventional curve analysis or pharmacokinetic model. The 3-D morphological features including shape features (compactness, margin, and ellipsoid fitting) and texture features (based on the grey level co-occurrence matrix) of the segmented tumor were obtained to characterize the lesion. One hundred and thirty-two biopsy-proven lesions (63 benign and 69 malignant) were used to evaluate the performance of the proposed computer-aided system for breast MRI. Five combined features including rate constant (k_ep), volume of plasma (v_p), energy (G₁), entropy (G₂), and compactness (C₁), had the best performance with an accuracy of 91.67% (121/132), sensitivity of 91.30% (63/69), specificity of 92.06% (58/63), and A_z value of 0.9427. Combining the kinetic and morphological features of 3-D breast MRI is a potentially useful and robust algorithm when attempting to differentiate benign and malignant lesions.     

Brain tissue classification based on DTI using an improved Fuzzy C-means algorithm with spatial constraints

We present an effective method for brain tissue classification based on diffusion tensor imaging (DTI) data. The method accounts for two main DTI segmentation obstacles: random noise and magnetic field inhomogeneities. In the proposed method, DTI parametric maps were used to resolve intensity inhomogeneities of brain tissue segmentation because they could provide complementary information for tissues and define accurate tissue maps. An improved fuzzy c-means with spatial constraints proposal was used to enhance the noise and artifact robustness of DTI segmentation. Fuzzy c-means clustering with spatial constraints (FCM_S) could effectively segment images corrupted by noise, outliers, and other imaging artifacts. Its effectiveness contributes not only to the introduction of fuzziness for belongingness of each pixel but also to the exploitation of spatial contextual information. We proposed an improved FCM_S applied on DTI parametric maps, which explores the mean and covariance of the feature spatial information for automated segmentation of DTI. The experiments on synthetic images and real-world datasets showed that our proposed algorithms, especially with new spatial constraints, were more effective.     

Dynamic sequential 3D gadolinium-enhanced MRI of the whole breast

Dynamic contrast-enhanced 2D MR imaging of the breast has shown high sensitivity and specificity for the detection and characterization of breast lesions. We investigated the ability of a dynamic fast 3D MR imaging technique that repeatedly scans the whole breast in 44-s intervals without an interscan delay time to obtain similar sensitivity and specificity as 2D imaging. Fifty-six patients scheduled for breast biopsy were entered into the study, and 83 lesions detected by 3D dynamic scanning were biopsied. Dynamic 3D contrast-enhanced breast imaging with subtraction detected and correctly classified all 23 cancers, and 44 of the 60 benign lesions yielding a sensitivity of 100%, a specificity of 73%, and a 100% predictive negative value. The enhancement profiles of metastatic lymph nodes were similar to those of primary cancer. This technique allowed detection of multifocal and multicentric lesions and did not require a priori knowledge of lesion location. These results indicate that dynamic contrast-enhanced 3D MRI of the whole breast is a useful and economically feasible method for staging breast cancer, providing a comprehensive noninvasive method for total evaluation of the breast and axilla in patients considering breast conservation surgery or lumpectomy.     

Automated vessel segmentation using cross-correlation and pooled covariance matrix analysis

Time-resolved contrast-enhanced magnetic resonance angiography (CE-MRA) provides contrast dynamics in the vasculature and allows vessel segmentation based on temporal correlation analysis. Here we present an automated vessel segmentation algorithm including automated generation of regions of interest (ROIs), cross-correlation and pooled sample covariance matrix analysis. The dynamic images are divided into multiple equal-sized regions. In each region, ROIs for artery, vein and background are generated using an iterative thresholding algorithm based on the contrast arrival time map and contrast enhancement map. Region-specific multi-feature cross-correlation analysis and pooled covariance matrix analysis are performed to calculate the Mahalanobis distances (MDs), which are used to automatically separate arteries from veins. This segmentation algorithm is applied to a dual-phase dynamic imaging acquisition scheme where low-resolution time-resolved images are acquired during the dynamic phase followed by high-frequency data acquisition at the steady-state phase. The segmented low-resolution arterial and venous images are then combined with the high-frequency data in k-space and inverse Fourier transformed to form the final segmented arterial and venous images. Results from volunteer and patient studies demonstrate the advantages of this automated vessel segmentation and dual phase data acquisition technique.     

A workflow for automated segmentation of the liver surface,hepatic vasculature and biliary tree anatomy from multiphase MR images

Accurate assessment of 3D models of patient-specific anatomy of the liver, including underlying hepatic and biliary tree, is critical for preparation and safe execution of complex liver resections, especially due to high variability of biliary and hepatic artery anatomies. Dynamic MRI with hepatospecific contrast agents is currently the only type of diagnostic imaging that provides all anatomical information required for generation of such a model, yet there is no information in the literature on how the complete 3D model can be generated automatically. In this work, a new automated segmentation workflow for extraction of patient-specific 3D model of the liver, hepatovascular and biliary anatomy from a single multiphase MRI acquisition is developed and quantitatively evaluated. The workflow incorporates course 4D k-means clustering estimation and geodesic active contour refinement of the liver boundary, based on organ’s characteristic uptake of gadolinium contrast agents overtime. Subsequently, hepatic vasculature and biliary ducts segmentations are performed using multiscale vesselness filters. The algorithm was evaluated using 15 test datasets of patients with liver malignancies of various histopathological types. It showed good correlation with expert manual segmentation, resulting in an average of 1.76 ± 2.44 mm Hausdorff distance for the liver boundary, and 0.58 ± 0.72 and 1.16 ± 1.98 mm between centrelines of biliary ducts and liver veins, respectively. A workflow for automatic segmentation of the liver, hepatic vasculature and biliary anatomy from a single diagnostic MRI acquisition was developed. This enables automated extraction of 3D models of patient-specific liver anatomy, and may facilitating better perception of organ’s anatomy during preparation and execution of liver surgeries. Additionally, it may help to reduce the incidence of intraoperative biliary duct damage due to an unanticipated variation in the anatomy.     

Breast disease evaluation with fat-suppressed magnetic resonance imaging.

Thirty patients with a variety of pathologically confirmed malignant and benign pathologic lesions of the breast were evaluated with a spectrally selective fat suppression imaging technique to obtain fat-suppressed images of the breast. The technique, a selective partial inversion-recovery (SPIR) method, demonstrated the architectural relationship of malignant and benign tumors with respect to the normal water-containing elements of the breast. These relationships included signs of advanced malignant disease such as tissue retraction, invasive growth, and multicentricity, which appeared on the fat-suppressed images. Fat-suppressed imaging provided useful information for assessing the breasts of both pre- and postmenopausal women, especially in the latter group, where fatty involution of the breast is common. Microcysts, which are normally not visualized by conventional methods, were demonstrated and associated with patients having confirmed fibrocystic disease of the breast. As expected, the SPIR technique did not improve the ability to distinguish between tissues having similar T1 and T2 relaxation time values, such as malignant tumors and normal breast parenchymal tissues. The technique was able to demonstrate that the intense lipid signal, known to be responsible for obscuring the borders of water-fat interfaces and small tumors, could be eliminated in a variety of pathological settings.     

A novel CNN algorithm for pathological complete response prediction using an I-SPY TRIAL breast MRI database

PurposeTo apply our convolutional neural network (CNN) algorithm to predict neoadjuvant chemotherapy (NAC) response using the I-SPY TRIAL breast MRI dataset.MethodsFrom the I-SPY TRIAL breast MRI database, 131 patients from 9 institutions were successfully downloaded for analysis. First post-contrast MRI images were used for 3D segmentation using 3D slicer. Our CNN was implemented entirely of 3 × 3 convolutional kernels and linear layers. The convolutional kernels consisted of 6 residual layers, totaling 12 convolutional layers. Dropout with a 0.5 keep probability and L2 normalization was utilized. Training was implemented by using the Adam optimizer. A 5-fold cross validation was used for performance evaluation. Software code was written in Python using the TensorFlow module on a Linux workstation with one NVidia Titan X GPU.ResultsOf 131 patients, 40 patients achieved pCR following NAC (group 1) and 91 patients did not achieve pCR following NAC (group 2). Diagnostic accuracy of our CNN two classification model distinguishing patients with pCR vs non-pCR was 72.5 (SD ± 8.4), with sensitivity 65.5% (SD ± 28.1) and specificity of 78.9% (SD ± 15.2). The area under a ROC Curve (AUC) was 0.72 (SD ± 0.08).ConclusionIt is feasible to use our CNN algorithm to predict NAC response in patients using a multi-institution dataset.     

基于图的加权核K均值的图像多尺度分割

提出改进的最小割(IMC)模型以避免分割出小的孤立点集,研究了改进的最小割模型与加权核K均值之间的等价关系,列举了几种常见的用于建立图割模型边权值的相似度函数,并分析了其对分割结果的影响.在此基础上.设计了一个摹于图的加权核K均值图像多尺度分割方法,该方法既避免了基于图割的图像分割中图谱的求解问题,又避免了加权核K均值方法中核矩阵的选取问题,同时实现了对图像多尺度的分割.通过对该方法进行抗噪性能的分析,以及在光学图像上对实验结果进行比较,验证了所提出方法的有效性.     

A deep-learning semantic segmentation approach to fully automated MRI-based left-ventricular deformation analysis in cardiotoxicity

Left-ventricular (LV) strain measurements with the Displacement Encoding with Stimulated Echoes (DENSE) MRI sequence provide accurate estimates of cardiotoxicity damage related to breast cancer chemotherapy. This study investigated an automated LV chamber quantification tool via segmentation with a supervised deep convolutional neural network (DCNN) before strain analysis with DENSE images. Segmentation for chamber quantification analysis was conducted with a custom DeepLabV3+ DCNN with ResNet-50 backbone on 42 female breast cancer datasets (22 training-sets, eight validation-sets and 12 independent test-sets). Parameters such as LV end-diastolic diameter (LVEDD) and ejection fraction (LVEF) were quantified, and myocardial strains analyzed with the Radial Point Interpolation Method (RPIM). Myocardial classification was validated against ground-truth with sensitivity-specificity analysis, the metrics of Dice, average perpendicular distance (APD) and Hausdorff-distance. Following segmentation, validation was conducted with the Cronbach's Alpha (C-Alpha) intraclass correlation coefficient between LV chamber quantification results with DENSE and Steady State Free Precession (SSFP) acquisitions and a vendor tool-based method to segment the DENSE data, and similarly for myocardial strain analysis in the chambers. The results of myocardial classification from segmentation of the DENSE data were accuracy = 97%, Dice = 0.89 and APD = 2.4 mm in the test-set. The C-Alpha correlations from comparing chamber quantification results between the segmented DENSE and SSFP data and vendor tool-based method were 0.97 for LVEF (56 ± 7% vs 55 ± 7% vs 55 ± 6%, p = 0.6) and 0.77 for LVEDD (4.6 ± 0.4 cm vs 4.5 ± 0.3 cm vs 4.5 ± 0.3 cm, p = 0.8). The validation metrics against ground-truth and equivalent parameters obtained from the SSFP segmentation and vendor tool-based comparisons show that the DCNN approach is applicable for automated LV chamber quantification and subsequent strain analysis in cardiotoxicity.