基于光学相干层析的视网膜图像分割

利用光学相干层析（optical coherence tomography,OCT）技术可以得到清晰的视网膜层状结构,实现视网膜层状结构自动分割功能是解决OCT技术应用于视网膜疾病诊断的一项基础问题。本文通过图像平滑、峰值探测、Snake模型、贪婪算法和样条插值等综合技术,对OCT视网膜图像进行分割,自动提取层状结构轮廓并获取视网膜厚度分布图。将以上算法应用于24例正常人眼底图像,并与专家手动标记轮廓提取的厚度相比,结果证实上述视网膜自动测量算法与专家人工标记取得较好一致性。本文提出的测量算法有望应用于视网膜变异性评估。     

Minimum distance mapping using three-dimensional optical coherence tomography for glaucoma diagnosis

Objective imaging of the optic nerve structure has become central to the management of patients with glaucoma. There is an urgent need in diagnosis and staging for reliable objective precursors and markers. Three-dimensional ultrahigh-resolution frequency domain optical coherence tomography (3D UHR OCT) holds particular promise in this respect since it enables volumetric assessment of intraretinal layers including tomographic data for the retinal nerve fiber layer (RNFL) and optic nerve head. The integrated analysis of this information and the resolution advantage has enabled the development of more informative indices of axonal damage in glaucoma compared with measurements of RNFL thickness and cup-to-disc ratio provided by commercial OCT devices. The potential for UHR OCT in enabling the combined analysis of tomographic and volumetric data on retinal structure is explored. A novel parameter was developed; the three-dimensional minimal distance as the optical correlate of true retinal nerve fiber layer thickness around the optic nerve head region. For the purposes of this pilot study, we present data from a normal subject and from two patients with characteristic optic nerve and retinal nerve fiber layer changes secondary to glaucoma.     

Quantitative analysis of the intraretinal layers and optic nerve head using ultra-high resolution optical coherence tomography

This study is designed to test the repeatability of the quantitative analysis of intraretinal layer thickness and cup-disc ratio of the optic nerve head using ultra-high resolution optical coherence tomography (UHR-OCT). Group A, containing 23 eyes of 12 healthy subjects, was imaged twice and group B, containing eight eyes of four subjects, was imaged three times. Intraretinal layers were segmented manually and the cup-to-disc ratio of the optic nerve head was analyzed. Custom-built automatic segmentation software was also used to segment a set of images for comparison. A total of nine intraretinal layers were visualized and extracted manually. With group A, the central foveal thickness was 186.4 ± 15.9 μm (mean ± SD). The average retinal thickness was 296.4 ± 21.3 μm. The best repeatability, obtained when two repeated scans were taken, was obtained for the outer nuclear layer followed by the ganglion cell layer, the inner nuclear layer, the retinal nerve fiber layer and the worst was obtained for the outer segment. The intraclass correlation ranged from 0.824 to 0.997. The coefficients of repeatability ranged from 3.24 to 18.3 μm, corresponding to 1.47% to 26.20%. With group B, high interclass correlations were found and the automatic segmentation results were compatible with the manual results. Our results indicated that more retinal features might be imageable using UHR-OCT.     

Enhancing the signal-to-noise ratio in ophthalmic optical coherence tomography by image registration--method and clinical examples

Optical coherence tomography (OCT) has already proven an important clinical tool for imaging and diagnosing retinal diseases. Concerning the standard commercial ophthalmic OCT systems, speckle noise is a limiting factor with respect to resolving relevant retinal features. We demonstrate successful suppression of speckle noise from mutually aligning a series of in vivo OCT recordings obtained from the same retinal target using the Stratus system from Humphrey-Zeiss. Our registration technique is able to account for the axial movements experienced during recording as well as small transverse movements of the scan line from one scan to the next. The algorithm is based on a regularized shortest path formulation for a directed graph on a map formed by interimage (B-scan) correlations. The resulting image enhancement typically increases the contrast-to-noise ratio (CNR) with a factor of three or more and facilitates segmentation and quantitative characterization of pathologies. The method is currently successfully being applied by medical doctors in a number of specific retinal case studies.     

Reduction of image artifacts in three-dimensional optical coherence tomography of skin in vivo

This paper presents results of in vivo studies on the effect of refractive index-matching media on image artifacts in optical coherence tomography (OCT) images of human skin. These artifacts present as streaks of artificially low backscatter and displacement or distortion of features. They are primarily caused by refraction and scattering of the OCT light beam at the skin surface. The impact of the application of glycerol and ultrasound gel is assessed on both novel skin-mimicking phantoms and in vivo human skin, including assessment of the epidermal thickening caused by the media. Based on our findings, recommendations are given for optimal OCT imaging of skin in vivo.     

基于联合决策卷积神经网络的光学相干断层扫描图像自动分类

光学相干断层扫描(OCT)技术能实现视网膜的高分辨率三维层析成像,对视网膜疾病类型的诊断和发展阶段的分析具有至关重要的作用。临床基于 OCT 图像的视网膜疾病诊断主要依靠眼科医生对图像中病变结构的分析,这一人工分析过程不仅耗时而且易产生主观的误判。研究视网膜疾病的自动分析和诊断技术将极大减轻眼科医生的工作量,是实现高效诊疗的有效途径。针对视网膜OCT图像自动分类,构建一种联合决策的卷积神经网络分类模型。该模型利用卷积神经网络从原始输入OCT图像中自动地学习不同层级的特征,同时在网络多个卷积层上设计多个决策层,这些决策层能够根据网络中不同尺度的特征图分别对OCT图像分类,最后模型融合所有决策层的分类结果做出最终决策。在Duke数据集(3 231张OCT图像)上的实验结果表明,基于多层级特征联合决策的卷积神经网络分类模型对正常视网膜、视网膜年龄相关性黄斑变性和视网膜黄斑水肿的平均识别准确率达到94.5%,灵敏性达到90.5%,特异性达到95.8%。在HUCM数据集(4 322张OCT图像)上的实验结果表明,基于多层级特征联合决策的卷积神经网络分类模型的平均识别准确率达到89.6%,灵敏性达到88.8%,特异性达到90.8%。充分利用卷积神经网络中丰富的多层级特征,能够有效地对视网膜OCT图像实现准确的分类,为临床上视网膜疾病的辅助诊断提供技术支撑。     

A novel kernelized fuzzy C-means algorithm with application in medical image segmentation

Image segmentation plays a crucial role in many medical imaging applications. In this paper, we present a novel algorithm for fuzzy segmentation of magnetic resonance imaging (MRI) data. The algorithm is realized by modifying the objective function in the conventional fuzzy C-means (FCM) algorithm using a kernel-induced distance metric and a spatial penalty on the membership functions. Firstly, the original Euclidean distance in the FCM is replaced by a kernel-induced distance, and thus the corresponding algorithm is derived and called as the kernelized fuzzy C-means (KFCM) algorithm, which is shown to be more robust than FCM. Then a spatial penalty is added to the objective function in KFCM to compensate for the intensity inhomogeneities of MR image and to allow the labeling of a pixel to be influenced by its neighbors in the image. The penalty term acts as a regularizer and has a coefficient ranging from zero to one. Experimental results on both synthetic and real MR images show that the proposed algorithms have better performance when noise and other artifacts are present than the standard algorithms.     

Analysis of posterior retinal layers in spectral optical coherence tomography images of the normal retina and retinal pathologies

We present a computationally efficient, semiautomated method for analysis of posterior retinal layers in three-dimensional (3-D) images obtained by spectral optical coherence tomography (SOCT). The method consists of two steps: segmentation of posterior retinal layers and analysis of their thickness and distance from an outer retinal contour (ORC), which is introduced to approximate the normal position of external interface of the healthy retinal pigment epithelium (RPE). The algorithm is shown to effectively segment posterior retina by classifying every pixel in the SOCT tomogram using the similarity of its surroundings to a reference set of model pixels from user-selected area(s). Operator intervention is required to assess the quality of segmentation. Thickness and distance maps from the segmented layers and their analysis are presented for healthy and pathological retinas.     

Ultrahigh-speed non-invasive widefield angiography

Retinal and choroidal vascular imaging is an important diagnostic benefit for ocular diseases such as age-related macular degeneration. The current gold standard for vessel visualization is fluorescence angiography. We present a potential non-invasive alternative to image blood vessels based on functional Fourier domain optical coherence tomography (OCT). For OCT to compete with the field of view and resolution of angiography while maintaining motion artifacts to a minimum, ultrahigh-speed imaging has to be introduced. We employ Fourier domain mode locking swept source technology that offers high quality imaging at an A-scan rate of up to 1.68 MHz. We present retinal angiogram over ～48 deg acquired in a few seconds in a single recording without the need of image stitching. OCT at 1060 nm allows for high penetration in the choroid and efficient separate characterization of the retinal and choroidal vascularization.     

Retinal response of Macaca mulatta to picosecond laser pulses of varying energy and spot size

We investigate the relationship between the laser beam at the retina (spot size) and the extent of retinal injury from single ultrashort laser pulses. From previous studies it is believed that the retinal effect of single 3-ps laser pulses should vary in extent and location, depending on the occurrence of laser-induced breakdown (LIB) at the site of laser delivery. Single 3-ps pulses of 580-nm laser energy are delivered over a range of spot sizes to the retina of Macaca mulatta. The retinal response is captured sequentially with optical coherence tomography (OCT). The in vivo OCT images and the extent of pathology on final microscopic sections of the laser site are compared. With delivery of a laser pulse with peak irradiance greater than that required for LIB, OCT and light micrographs demonstrate inner retinal injury with many intraretinal and/or vitreous hemorrhages. In contrast, broad outer retinal injury with minimal to no choriocapillaris effect is seen after delivery of laser pulses to a larger retinal area (60 to 300 microm diam) when peak irradiance is less than that required for LIB. The broader lesions extend into the inner retina when higher energy delivery produces intraretinal injury. Microscopic examination of stained fixed tissues provide better resolution of retinal morphology than OCT. OCT provides less resolution but could be guided over an in vivo, visible retinal lesion for repeated sampling over time during the evolution of the lesion formation. For 3-ps visible wavelength laser pulses, varying the spot size and laser energy directly affects the extent of retinal injury. This again is believed to be partly due to the onset of LIB, as seen in previous studies. Spot-size dependence should be considered when comparing studies of retinal effects or when pursuing a specific retinal effect from ultrashort laser pulses.     

Active retinal tracker for clinical optical coherence tomography systems

An active, hardware-based retinal tracker is integrated with a clinical optical coherence tomography (OCT) system to investigate the effects of stabilization on acquisition of high-resolution retinal sections. The prototype retinal tracker locks onto common fundus features, detects transverse eye motion via changes in feature reflectance, and positions the OCT diagnostic beam to fixed coordinates on the retina with mirrors driven by a feedback control loop. The system is tested in a full clinical protocol on subjects with normal and glaucomatous eyes. Experimental analysis software is developed to coalign and coadd multiple fundus and OCT images and to extract quantitative information on the location of structures in the images. Tracking is highly accurate and reproducible on all but one subject, resulting in the ability to scan the same retinal location continually over long periods of time. The results show qualitative improvement in 97% of coadded OCT scans and a reduction in the variance of the position of the optic disc cup edge to less than 1 pixel (< 60 microm). The tracking system can be easily configured for use in research on ultra-high-resolution OCT systems for advanced image modalities. For example, tracking will enable very high density 3-D scans of the retina, which are susceptible to eye motion artifacts even for new high-speed systems.     

基于卷积神经网络和图像显著性的心脏CT图像分割

目的心脏医学影像中,感兴趣部分的提取与分割是诊断心脏病变部位的关键。由于心脏舒张、收缩以及血液的流动,心脏CT图像易出现弱边界、伪影,传统分割算法易产生过度分割的情况。为此,提出一种基于卷积神经网络和图像显著性的心脏CT图像分割方法。方法采用卷积神经网络对目标区域进行定位,滤除肋骨、肌肉等造影对比不明显部分,截取出感兴趣区域,结合感兴趣区域的对比度计算并提高感兴趣区域的心脏组织的显著值。通过获得的显著值图像截取心脏图像,并与区域生长算法的分割结果进行对比。最后使用泰州人民医院11例患者的影像数据对算法模型进行训练和测试,随机选择9例用于训练,剩余2例用于测试。结果所提算法模型在心底、心中、心尖3个心脏分段的分割正确率分别达到了92.79%、92.79%、94.11%,均优于基于区域生长的分割方法。结论基于卷积神经网络和图像显著性的分割方法能够准确获取心脏的外围轮廓,轮廓边缘更加平滑,完全能够满足CT图像序列的心脏全自动分割任务需求,分割后的图像更有利于医生对患者心脏健康状况和病变部位的观察。     

Automated quantification of three-dimensional subject motion to monitor image quality in high-resolution peripheral quantitative computed tomography

Subject motion during acquisition of high-resolution peripheral quantitative computed tomography (HR-pQCT) results in image artifacts and interferes with quantification of bone architecture used to study bone-related diseases such as osteoporosis. We propose an automatic method to measure physical subject motion that frequently takes place during acquisition. Three measures derived from projection data are proposed to quantify motion artifacts: in-plane translation (ε(T)) and in-plane rotation (ε(R)) utilizing projection moments and longitudinal translation (ε(z)) based on tracking projection profiles. Validation was performed using a phantom containing sections of distal human cadaver radii attached to a mechanical device to precisely control in-plane rotation and longitudinal translation that was intentionally performed during HR-pQCT data acquisition. Motion measured by the new automated technique was compared to the known applied motion, and related to percent errors in morphological parameters quantifying bone properties. It was determined that of the three proposed measures, ε(T) best captured a quantified representation of image quality. ε(T) linearly relates to true physical in-plane translational motion (r(2) = 0.95, p<0.001) and is independent from longitudinal translational motion as well as the object being scanned. Additionally, ε(z) captures large longitudinal movements and combines well with ε(T) to fully characterize physical motion artifacts. The magnitude of ε(T) corresponds to morphological parameter error and is an excellent basis to select high-quality images. Morphological parameter errors from these experiments confirmed our earlier computer simulations which showed that increased subject motion resulted in artificially higher trabecular number, and artificially lower bone mineral density and cortical thickness. The magnitude and, notably, the uncertainty of the morphological errors increased with increased physical motion, and this impedes a direct linear compensation of parameter errors. The automated method presented provides a basis for consistent and objective quality assurance for HR-pQCT scanning, and addresses an important challenge for this novel imaging modality that is rapidly becoming an important basis for assessment and monitoring of bone quality.     

Segmentation and quantification of retinal lesions in age-related macular degeneration using polarization-sensitive optical coherence tomography

We present polarization-sensitive optical coherence tomography (PS-OCT) for quantitative assessment of retinal pathologies in age-related macular degeneration (AMD). On the basis of the polarization scrambling characteristics of the retinal pigment epithelium, novel segmentation algorithms were developed that allow one to segment pathologic features such as drusen and atrophic zones in dry AMD as well as to determine their dimensions. Results from measurements in the eyes of AMD patients prove the ability of PS-OCT for quantitative imaging based on the retinal features polarizing properties. Repeatability measurements were performed in retinas diagnosed with drusen and geographic atrophy in order to evaluate the performance of the described methods. PS-OCT appears as a promising imaging modality for three-dimensional retinal imaging and ranging with additional contrast based on the structures' tissue-inherent polarization properties.     

Fetal ultrasound image segmentation system and its use in fetal weight estimation

A semi-automated fetal ultrasound image segmentation system is developed to improve the estimation of fetal weight (EFW). Four standardized fetal parameters are measured by the proposed segmentation system: biparietal diameter, head circumference, abdominal circumference and femur length. Computerized measurements of 215 fetuses are compared with manual measurements in term of fitness analysis and difference analysis. Among 215 cases, computerized measurements of 103 fetuses within 3 days of delivery are utilized in the fetal weight estimation. The EFW based on computerized measurements and manual measurements are compared by using regression analysis, artificial neural network and support vector regression. By using different estimation methods, the computerized measurements decrease the EFW errors about 40–70 g. The lowest mean absolute percentage error of EFW decrease from 6.71% for manual measurements to 4.66% for computerized measurements. The proposed fetal ultrasound image segmentation system can provide more accurate EFW in antepartum examination.     

一种基于组合分割思想的尿沉渣图像分割新方法

本研究针对复杂散焦的尿沉渣图像的精细分割,提出了首先使用小波变换和形态学处理消除散焦影响并进行图像的粗分割,然后根据粗分割得到的子图像的情况采用边缘检测或者自适应阈值分割的混合分割方法进行细分割,最后再采用剥离算法处理待分割的粘连重叠成分的分割。该方法不受散焦影响,充分利用了图像的多种信息,因此分割结果准确。实验结果表明,该方法对尿沉渣图像的分割有效且令人满意。     

Ultrasound image matching using genetic algorithms

Image compounding reduces the artifacts inherent in ultrasound imaging, but accurate matching of images for compounding depends on their accurate placement in the compound image plane. A method is presented to reduce displacement errors during compounding of ultrasound B-scans of a normal human shank. A genetic algorithm was used to place matching B-scans in the compound image. The method was tested on a phantom and was shown to reduce, but not eliminate, mismatches due to the displacement of B-scans from their original position in the compound image plane. The results can be extended to applications in lower-limb prosthetics, where ultrasound imaging can be used to visualise the internal geometry of amputees' residual limbs.     

Non-invasive determination of the absorption coefficient of the brain from time-resolved reflectance using a neural network

We investigated the performance of a neural network for derivation of the absorption coefficient of the brain from simulated non-invasive time-resolved reflectance measurements on the head. A five-layered geometry was considered assuming that the optical properties (except the absorption coefficient of the brain) and the thickness of all layers were known with an uncertainty. A solution of the layered diffusion equation was used to train the neural network. We determined the absorption coefficient of the brain with an RMS error of <6% from reflectance data at a single distance calculated by diffusion theory. By applying the neural network to reflectance curves obtained from Monte Carlo simulations, similar errors were found.     

Validation of optical coherence tomography in vivo using cryostat histology

We aimed to validate for the first time optical coherence tomography (OCT) measurements of epidermal thickness (ET) using cryopreparation for histology. OCT assessments of ET were performed on healthy skin using the algorithms as follows: first, peak-to-valley analysis of the A-scan (ET-OCT-V), second, line-traced image analysis of the B-scan (ET-OCT-IA). Histology was performed using cryostat sections which were also evaluated using the image analysis (ET-Histo). We selected 114 samples, including B-scans and corresponding histology, for method comparison between ET-OCT-IA and ET-Histo. Forty-two A-scans were available for method comparison between ET-OCT-V and ET-Histo. Bland and Altman plots revealed a marked bias with wide 95% limits of agreement for ET-OCT-V versus ET-Histo. Comparison of ET-OCT-IA versus ET-Histo revealed only a slight bias and narrow 95% limits of agreement. A-scan analysis for ET determination is linked to significant limitations and lacks agreement with histology. By contrast, we observed satisfactory agreement between ET-OCT-IA and ET-Histo indicating that both methods can be utilized interchangeably. OCT using the line-traced image analysis of the B-scan appears to be a valid and relatively practicable method for the determination of ET in vivo. Furthermore, the comparisons with the in vivo OCT profiles demonstrate that cryostat sectioning provides a better preservation of relative and absolute dimensions of skin layers than paraffin embedding.