结合先验稀疏字典和空洞填充的CT图像肝脏分割

针对复杂多变的肝脏图像,提出了一种基于先验稀疏字典和空洞填充的三维肝脏图像分割方法。对腹部CT图像进行Gabor特征提取,并分别在Gabor图像和灰度图像的肝脏金标准边界上选择大小相同的图像块作为两组训练集,利用训练集得到两种查询字典及稀疏编码。将金标准图像与待分割图像配准,并将配准后的肝脏边界作为待分割图像的肝脏初始边界;在初始边界点上的十邻域内选择大小相同的两组图像块作为测试样本,利用测试样本与查询字典计算稀疏编码及重构误差,并选择重构误差最小的图像块的中心作为待分割肝脏的边界点;最后,设计一种空洞填充方法对肝脏边界进行补全和平滑处理,得到最终分割结果。利用医学图像计算和计算机辅助介入国际会议中提供的肝脏数据进行了实验验证。结果表明,该方法对肝脏分割图像具有较好的适用性和鲁棒性,并获得了较高的分割精度。其中,平均体积重叠率误差为(5.21±0.45)%,平均相对体积误差为(0.72±0.12)%,平均对称表面距离误差为(0.93±0.14)mm。

原木横截面CT图像的伪彩色增强方法对比研究

CT灰度图像具有分辨率低,不利于人眼直观分析等不足,而CT图像直观分析是最基本的CT分析方法。为提高原木CT图像的分辨率,利用在不同温度(15、-5、-20℃)下扫描获取的杉木、红松和冷杉三个树种原木横截面CT图像,采用灰度级-彩色变换方法,通过彩虹码和热金属码的编码方法实现图像伪彩色增强,并对两种方法在不同温度下CT图像的增强效果和特征进行对比分析。研究表明:(1)伪彩色增强方法能显著提高CT图像的视觉分辨率,减少直观判断误差。如视觉灰度相近的A与B两处(CT值范围分别为-550^-680Hu和-530^-730Hu)在不同方法的增强图像中呈现出不同的颜色差异。(2)热金属码的视觉增强效果低于彩虹码。彩虹码能为CT图像直观分析提供更多的细观物理信息,而热金属码能更加明确、清晰地划分出木材内部不同密度的结构分布。(3)相较于CT灰度图像,伪彩色增强后的图像有利于研究不同温度下木材内部结构、各相组分的变化情况,为定量研究温度对木材内部水分的状态和含量变化提供新的思路。

基于多尺度并行深度可拆分的CNN新冠肺炎CT图像去噪方法

目前新冠肺炎(COVID-19)在全球蔓延,为了对新冠肺炎进行早期诊断,同时减轻医护人员的工作压力,使用深度学习对患者胸部电子计算机断层扫描(CT)图像进行分析变得越来越重要。针对肺炎图像中纹理细节较为丰富、边缘结构模糊、极易干扰机器及医生诊断的问题,本文提出一种基于多尺度并行深度可拆分卷积神经网络(MSP-ReCNN),对新冠肺炎CT图像进行去噪处理,提升肺炎图像质量。多尺度特征提取模块从不同尺度提取肺炎图像中的纹理特征细节,采用深浅通道并行方式,分别提取肺炎图像中的高维度以及低维度的特征。为进一步优化网络模型,提出一种拆分卷积方式,可将特征图拆分为两类,一类为主要关注特征,另一类为次要关注特征。使用复杂度高的计算方式从主要关注特征中提取关键信息,对于次要关注特征,则采取复杂度低的计算方式提取补偿信息。通过与非局部均值(NLM)去噪算法、收缩卷积神经网络(SCNN)深度模型、去噪卷积神经网络(DnCNN)深度模型对比,以及网络消融实验,可以看出本文提出的模型能有效去除肺炎图像中的噪声,并且可以更好地保留原始图像中的纹理结构细节,为机器以及医生提供更可靠的辅助诊断。

基于GAN的岩石CT图像超分辨算法

岩石电子计算机断层扫描(Computed Tomography,CT)图像可使岩石内部结构可视化,故可以作为评估油气储藏量的依据。但是,受采集条件的约束,岩石CT图像往往分辨率低、细节较模糊。因此,本文基于生成对抗网络(Generative Adversarial Network,GAN)方法引入岩石孔隙度损失,在原有损失函数约束的基础上增加新的约束项,并进行4倍超分辨重构实验,通过峰值信噪比(Peak Signal to Noise Ratio,PSNR)和结构相似性(Structural Similarity Index,SSIM)的指标比较重建结果。结果表明,本文方法在指标和视觉上均取得了良好效果。

Primary clear cell carcinoma in the liver: CT and MRI findings

AIM: To retrospectively analyze the computed tomography (CT) and magnetic resonance imaging (MRI) appearances of primary clear cell carcinoma of the liver (PCCCL) and compare the imaging appearances of PCCCL and common type hepatocellular carcinoma (CHCC) to determine whether any differences exist between the two groups. METHODS: Twenty cases with pathologically proven PCCCL and 127 cases with CHCC in the Second Affiliated Hospital of Sun Yat-sen University were included in this study. CT or MRI images from these patients were retrospectively analyzed. The following imaging findings were reviewed: the presence of liver cirrhosis, tumor size, the enhancement pattern on dynamic contrast scanning, the presence of pseudo capsules, tumor rupture, portal vein thrombosis and lymph node metastasis. RESULTS: Both PCCCL and CHCC were prone to occur in patients with liver cirrhosis, the association rate of liver cirrhosis was 80.0% and 78.7%, respectively (P > 0.05). The mean sizes of PCCCL and CHCC tumors were (7.28 ± 4.25) cm and (6.96 ± 3.98) cm, respectively. Small HCCs were found in 25.0% (5/20) of PCCCL and 19.7% (25/127) of CHCC cases. No significant differences in mean size and ratio of small HCCs were found between the two groups (P = 0.658 and 0.803, respectively). Compared with CHCC patients, PCCCL patients were more prone to form pseudo capsules (49.6% vs 75.0%, P = 0.034). Tumor rupture, typical HCC enhancement patterns and portal vein tumor thrombosis were detected in 15.0% (3/20), 72.2% (13/18) and 20.0% (4/20) of patients with PCCCL and 3.1% (4/127), 83.6% (97/116) and 17.3% (22/127) of patients with CHCC, respectively. There were no significant differences between the two groups (all P > 0.05). No patients with PCCCL and 2.4% (3/127) of patients with CHCC showed signs of lymph node metastasis (P > 0.05). CONCLUSION: The imaging characteristics of PCCCL are similar to those of CHCC and could be useful for differentiating these from other liver tumors (such as hemangioma and hepatic metastases). PCCCLs are more prone than CHCCs to form pseudo capsules.

The optimal slice thickness of CT in revealing lobulation of malignant solitary pulmonary nodules

Objective： The aim of this study was to determine an optimal slice thickness that was efficient in revealing Iobulation of malignant solitary pulmonary nodules （SPNs） on multi-slice spiral computed tomography （MSCT） images preliminarily. Methods： Fifty patients with malignant SPNs （diameter -〈 3 cm） underwent multidetector-row computed tomography of the chest in a single-breath-hold technique. The raw data were acquired with a collimation of 0.625 mm. Three sets of contiguous images were reconstructed with 1-, 2-, and 5-ram slice thickness, respectively. The Iobulation sign of SPNs on the computed tomography （CT） images presented in 1-, 2-, and 5-ram slice thickness was compared. Using the 1-mm sections as the gold standard, an optimal slice thickness in revealing Iobulation sign of SPNs was determined. Results： The 1-mm-thick images CT revealed 98 Iobulations （25 with chord distance 〈 1 ram; 30 with chord distance 1-2 ram; 43 with chord distance 〉 2 mm） of 45 malignant SPNs. 18 Iobulations with chord distance 〈 1 mm presented in 2-mm-thick sections were as same as those in I-ram-thick sections. Statistically significant difference in Iobulations number was found between that revealed in 2-ram-thick images and that in I-ram-thick images （P = 0.023 〈 0.05）. 16 Iobulations with chord distance 〈 1 mm presented in 5-mm-thick sections were as same as that in I-ram-thick sections. There was statistically significant difference in Iobulations number between that revealed in 5-mm-thick images and that in I-ram-thick images （P = 0.004 〈 0.05）. The 24 Iobulations with chord distance 1-2 mm presented in 2-ram-thick sections were as same as that in 1-mm-thick sections. No statistically significant difference in Iobulations number were found between that revealed in 2-mm-thick images and that in 1-mm-thick images （P = 0.261 〉 0.05）. 13 Iobulations with chord distance 1-2 mm presented in 5-ram-thick sections were as same as that in 1- mm-thick sections. There was statistically significant difference in Iobulations number between that revealed in 5-ram-thick images and that in I-ram-thick images （P = 0.003 〈 0.05）. 40 Iobulations with chord distance 〉 2 mm presented in 2-ram-thick sections were as same as that in I-ram-thick sections. No statistically significant difference in Iobulations number was found between that revealed in 2-ram-thick images and that in I-ram-thick images （P = 0.631 〉 0.05）. 36 Iobulations with chord distance 〉 2 mm presented in 5-mm-thick sections were as same as that in I-ram-thick sections. There was no statistically significant difference in Iobulations number between that revealed in 5-ram-thick images and that in I-ram-thick images （P = 0.264 〉 0.05）. Conclusion： It is suggested that the use of 1-mm slice thickness is suitable in revealing Iobulations with chord distance 〈 1 ram. A 2-mm slice thickness is suggested to be used in revealing Iobulafions with chord distance 1-2 mm and 5-mm slice thickness to be used in revealing Iobulations with chord distance 〉 2 mm.

Pulmonary nodules: optimal slice thickness of CT in revealing bronchial imageology

Objective： The aim of our study was to determine an optimal slice thickness that was efficient in revealing bronchial imageology of pulmonary nodules （PNs） on multi-slice spiral computed tomography （MSCT） images preliminarily. Methods： Fifty-four patients with 62 PNs （diameter 〈 3 cm） underwent multidetector-row computed tomography of the chest in a single-breath-hold technique. The raw data were acquired with a collimation of 0.625 mm. Three sets of contiguous im- ages were reconstructed with 1-, 2-, and 5-mm slice thickness, respectively. Bronchial imageology of SPNs on the CT images presented in 1-, 2-, and 5-mm slice thickness was compared. Using the 1-mm sections as the gold standard, an optimal slice thickness in revealing bronchial imageology of PNs was determined. Results： Bronchial imageology of PNs on the CT images presented in 1 mm slice thickness involved 85 bronchi （35 second-fourth generation bronchi; 50 fifth-eighth generation bronchi）. Bronchial imageology on 2-mm-thick sections was as same as that on I-mm-thick sections in 34 second- fourth generation bronchi. No statistically significant difference in number of second- fourth generation bronchi with same bronchial imageology was found between that on 2-mm-thick images and I-mm-thick images （P = 0.836 〉 0.05）. Bronchial imageology on 5-mm-thick sections was as same as that on 1-mm-thick sections in 24 second-fourth generation bronchi. There was statistically significant difference in number of second-fourth generation bronchi with same bronchial imageology between that on 5-mm-thick images and 1-mm-thick images （P = 0.026 〈 0.05）. Bronchial imageology on 2-mm-thick sections was as same as that on 1-mm-thick sections in 38 fifth-eighth generation bronchi. There was statistically significant difference in number of fifth-eighth generation bronchi with same bronchial imageology between that on 2-mm-thick images and 1-mm- thick images （P = 0.029 〈 0.05）. Bronchial imageology on 5-mm-thick images was as same as that on I-mm-thick images in 31 fifth-eighth generation bronchi. There was statistically significant difference in number of fifth-eighth generation bronchi with same bronchial imageology between that on 5-mm-thick sections and I-mm-thick sections （P = 0.001 〈 0.05）. Conclusion： It is suggested that the use of 2-mm slice thickness is suitable in revealing second- fourth generation bronchi and the use of l-mm slice thickness is suitable in revealing fifth-eighth generation bronchi.

Colorectal cancer and 18FDG-PET/CT:What about adding the T to the N parameter in loco-regional staging?

AIM:To evaluate whether FDG-positron emission tomography(PET)/computed tomography(CT)may be an accurate technique in the assessment of the T stage in patients with colorectal cancer.METHODS:Thirty four consecutive patients(20 men and 14 women;mean age:63 years)with a histologically proven diagnosis of colorectal adenocarcinoma and scheduled for surgery in our hospital were enrolled in this study.All patients underwent FDG-PET/CT preoperatively.The primary tumor site and extent were evaluated on PET/CT images.Colorectal wall invasion was analysed according to a modified T classification that considers only three stages(≤T2,T3,T4).Assessment of accuracy was carried out using 95%confidence intervals for T.RESULTS:Thirty five/37(94.6%)adenocarcinomas were identified and correctly located on PET/CT images.PET/CT correctly staged the T of 33/35 lesions identified showing an accuracy of 94.3%(95%CI:87%-100%).All T1,T3 and T4 lesions were correctly staged,while two T2 neoplasms were overstated as T3.CONCLUSION:Our data suggest that FDG-PET/CT may be an accurate modality for identifying primary tumor and defining its local extent in patients with colorectal cancer.

Synchronous double cancers of the common bile duct

We report an extremely rare case of synchronous double cancers of the common bile duct without pancreaticobiliary maljunction.Only two similar cases have been reported in the English literature.Endoscopic retrograde cholangiopancreatography showed a tuberous filling defect in the middle and superior parts of the common bile duct,and mild stenosis in the inferior duct.Computed tomography(CT) showed a well enhanced mass in the middle and superior parts of the common bile duct.A single cancer of the middle and superior bile duct was suspected and extra-hepatic bile duct resection was performed.CT eleven months after surgery revealed enhanced inferior bile duct wall and a slightly enhanced tumor within it.Retrospective review of the CT images taken before first surgery showed enhanced inferior bile duct wall without intrabiliary tumor only on the delayed phase.The inferior bile duct tumor was suspected to have originally co-existed with the middle and superior bile duct tumor.Pancreaticoduodenectomy was performed subsequently.Histopathological examination revealed that the middle and superior bile duct tumor was a moderately differentiated tubular adenocarcinoma while the inferior bile duct tumor was a papillary adenocarcinoma.The two tumors were separated and had different histological findings and growth patterns,further suggesting that they were two primary cancers.

Visualization of Three-dimensional Human Data Based on CT Image

Three-dimensional medical image visualization becomes an essential part for medical field, including computer aided diagnosis, surgery planning and simulation, artificial limb surgery, radiotherapy planning, and teaching etc. In this paper, marching cubes algorithm is adopted to reconstruct the 3-D images for the CT image sequence in DICOM format under theVC＋＋6.0 and the visual package VTK platform. The relatively simple interactive operations such as rotation and transfer can be realized on the platform. Moreover, the normal vector and interior point are calculated to form the virtual clipping plane, which is then used to incise the 3-D object. Information of the virtual slice can be obtained, in the mean while the virtual slice images are displayed on the screen. The technique can realize the real time interaction extraction of virtual slice on 3-D CT image. The cuboids structured can be zoomed, moved and eircumrotated by operating mouse to incise the 3-D reconstruction object. Real time interaction can be realized by clipping the reconstruction object. The coordinates can be acquired by the mouse clicking in the 3D space, to realize the point mouse pick-up as well angle and distance interactive measurement. We can get quantitative information about 3-D images through measurement.

涡轮叶片CT图像边缘提取的最优算子研究

为了解决叶片计算机断层(computed tomograghy,简称CT)图像边缘提取精度低效率差的问题,以解析法提取边缘的理论为基础,推导了适用于CT图像边缘提取的最优算子,并通过定量计算与模拟标准图像对最优算子的性能进行了分析,结果表明最优算子在抗噪、定位精度以及单像素响应等方面都比传统的边缘提取算子更为优越.实际应用最优算子提取的叶片边缘更光顺,噪声点以及伪边缘更少,进一步证明了最优算子在CT图像边缘提取方面有着明显的优势.

Contrast-enhanced multiple-phase imaging features in hepatic epithelioid hemangioendothelioma

AIM： To investigate and review the contrast-enhanced multiple-phase computed tomography （CEMP CT） and magnetic resonance imaging （MRI） findings in patients with pathologically confirmed hepatic epithelioid hemangioendothelioma （HEHE）. METHODS： Findings from imaging examinations in 8 patients （5 women and 3 men） with pathologically confirmed HEHE were retrospectively reviewed （CT images obtained from 7 patients and MR images obtained from 6 patients）. The age of presentation varied from 27 years to 60 years （average age 39.8 years）. RESULTS： There were two types of HEHE： multifocal type （n = 7） and diffuse type （n = 1）. Tn the multifocal-type cases, there were 74 lesions on CT and 28 lesions on MRI with 7 lesions found with diffusion weighted imaging; 18 （24.3%） of 74 lesions on plain CT and 26 （92.9%） of 28 lesions on pre-contrast MRI showed the target sign. On CEMP CT, 28 （37.8%） of 74 lesions appeared with the target sign and a progressive-enhancement rim and 9 （12.2%） of 74 lesions displayed progressive enhancement, maintaining a state of persistent enhancement. On CEMP MRI, 27 （96.4%） of 28 lesions appeared with the target sign with a progressive-enhancement rim and 28 （100%） of 28 lesions displayed progressive-enhancement, maintaining a state of persistent enhancement. In the diffuse-type cases, an enlarged liver was observed with a large nodule appearing with persistent enhancement on CEMP CT and MRI. CONCLUSION： The most important imaging features of HEHE are the target sign and/or progressive en- hancement with persistent enhancement on CEMP CT and MRI. MRI is advantageous over CT in displaying these imaging features.

Classification of Medical Brain Images

Since brain tumors endanger people’s living quality and even their lives, the accuracy of classification becomes more important. Conventional classifying techniques are used to deal with those datasets with characters and numbers. It is difficult, however, to apply them to datasets that include brain images and medical history (alphanumeric data), especially to guarantee the accuracy. For these datasets, this paper combines the knowledge of medical field and improves the traditional decision tree. The new classification algorithm with the direction of the medical knowledge not only adds the interaction with the doctors, but also enhances the quality of classification. The algorithm has been used on real brain CT images and a precious rule has been gained from the experiments. This paper shows that the algorithm works well for real CT data.

Abdominal aortic aneurysm： Treatment options, image visualizations and follow-up procedures

Abdominal aortic aneurysm is a common vascular disease that affects elderly population.Open surgical repair is regarded as the gold standard technique for treatment of abdominal aortic aneurysm,however,endovaseular aneurysm repair has rapidly expanded since its first introduction in 1990s.As a less invasive technique,endovascular aneurysm repair has been confirmed to be an effective alternative to open surgical repair,especially in patients with co-morbid conditions.Computed tomography （CT） angiography is currently the preferred imaging modality for both preoperative planning and post-operative follow-up.2D CT images are complemented by a number of 3D reconstructions which enhance the diagnostic applications of CT angiography in both planning and follow-up of endovascular repair.CT has the disadvantage of high cummulative radiation dose,of particular concern in younger patients,since patients require regular imaging follow-ups after endovascular repair,thus,exposing patients to repeated radiation exposure for life.There is a trend to change from CT to ultrasound surveillance of endovascular aneurysm repair.Medical image visualizations demonstrate excellent morphological assessment of aneurysm and stent-grafts,but fail to provide hemodynamic changes caused by the complex stent-graft device that is implanted into the aorta.This article reviews the treatment options of abdominal aortic aneurysm,various image visualization tools,and follow-up procedures with use of different modalities including both imaging and computational fluid dynamics methods.Future directions to improve treatment outcomes in the follow-up of endovascular aneurysm repair are outlined.

深度学习多部位病灶检测与分割

目的多部位病灶具有大小各异和类型多样的特点,对其准确检测和分割具有一定的难度。为此,本文设计了一种2.5D深度卷积神经网络模型,实现对多种病灶类型的计算机断层扫描(computed tomography,CT)图像的病灶检测与分割。方法利用密集卷积网络和双向特征金字塔网络组成的骨干网络提取图像中的多尺度和多维度信息,输入为带有标注的中央切片和提供空间信息的相邻切片共同组合而成的CT切片组。将融合空间信息的特征图送入区域建议网络并生成候选区域样本,再由多阈值级联网络组成的Cascade R-CNN(region convolutional neural networks)筛选高质量样本送入检测与分割分支进行训练。结果本文模型在DeepLesion数据集上进行验证。结果表明,在测试集上的平均检测精度为83.15%,分割预测结果与真实标签的端点平均距离误差为1.27 mm,直径平均误差为1.69 mm,分割性能优于MULAN(multitask universal lesion analysis network for joint lesion detection,tagging and segmentation)和Auto RECIST(response evaluation criteria in solid tumors),且推断每幅图像平均时间花费仅91.7 ms。结论对于多种部位的CT图像,本文模型取得良好的检测与分割性能,并且预测时间花费较少,适用病变类别与DeepLesion数据集类似的CT图像实现病灶检测与分割。本文模型在一定程度上能满足医疗人员利用计算机分析多部位CT图像的需求。