Automatic Correspondence on Medical Images: A Comparative Study of Four Methods for Allocating Corresponding Points

The accurate estimation of point correspondences is often required in a wide variety of medical image-processing applications. Numerous point correspondence methods have been proposed in this field, each exhibiting its own characteristics, strengths, and weaknesses. This paper presents a comprehensive comparison of four automatic methods for allocating corresponding points, namely the template-matching technique, the iterative closest points approach, the correspondence by sensitivity to movement scheme, and the self-organizing maps algorithm. Initially, the four correspondence methods are described focusing on their distinct characteristics and their parameter selection for common comparisons. The performance of the four methods is then qualitatively and quantitatively compared over a total of 132 two-dimensional image pairs divided into eight sets. The sets comprise of pairs of images obtained using controlled geometry protocols (affine and sinusoidal transforms) and pairs of images subject to unknown transformations. The four methods are statistically evaluated pairwise on all image pairs and individually in terms of specific features of merit based on the correspondence accuracy as well as the registration accuracy. After assessing these evaluation criteria for each method, it was deduced that the self-organizing maps approach outperformed in most cases the other three methods in comparison.     

Pulmonary nodule registration in serial CT scans based on rib anatomy and nodule template matching

An automated method is being developed in order to identify corresponding nodules in serial thoracic CT scans for interval change analysis. The method uses the rib centerlines as the reference for initial nodule registration. A spatially adaptive rib segmentation method first locates the regions where the ribs join the spine, which define the starting locations for rib tracking. Each rib is tracked and locally segmented by expectation-maximization. The ribs are automatically labeled, and the centerlines are estimated using skeletonization. For a given nodule in the source scan, the closest three ribs are identified. A three-dimensional (3D) rigid affine transformation guided by simplex optimization aligns the centerlines of each of the three rib pairs in the source and target CT volumes. Automatically defined control points along the centerlines of the three ribs in the source scan and the registered ribs in the target scan are used to guide an initial registration using a second 3D rigid affine transformation. A search volume of interest (VOI) is then located in the target scan. Nodule candidate locations within the search VOI are identified as regions with high Hessian responses. The initial registration is refined by searching for the maximum cross-correlation between the nodule template from the source scan and the candidate locations. The method was evaluated on 48 CT scans from 20 patients. Experienced radiologists identified 101 pairs of corresponding nodules. Three metrics were used for performance evaluation. The first metric was the Euclidean distance between the nodule centers identified by the radiologist and the computer registration, the second metric was a volume overlap measure between the nodule VOIs identified by the radiologist and the computer registration, and the third metric was the hit rate, which measures the fraction of nodules whose centroid computed by the computer registration in the target scan falls within the VOI identified by the radiologist. The average Euclidean distance error was 2.7 +/- 3.3 mm. Only two pairs had an error larger than 10 mm. The average volume overlap measure was 0.71 +/- 0.24. Eighty-three of the 101 pairs had ratios larger than 0.5, and only two pairs had no overlap. The final hit rate was 93/101.     

Evaluation of cost functions for gray value matching of two-dimensional images in radiotherapy

In external beam radiotherapy, portal imaging is applied for verification of the patient setup. Current automatic methods for portal image registration, which are often based on segmentation of anatomical structures, are especially successful for images of the pelvic region. For portal images of more complicated anatomical structures, e.g., lung, these techniques are less successful. It is desirable to have a method for image registration that is applicable for a wide range of treatment sites. In this study, a registration method for two-dimensional (2D) registration of portal and reference images based on intensity values was tested on portal images of various anatomical sites. Tests were performed with and without preprocessing (unsharp mask filtering followed by histogram equalization) for 96 image pairs and six cost functions. The images were obtained from treatments of the rectum, salivary gland, brain, prostate, and lung. To get insight into the behavior of the various cost functions, cost function values were computed for each portal image for 20,000 transformations of the corresponding reference image, translating the reference image in a range of +/- 1 cm and rotating +/- 10 degrees with respect to the clinical match. The automatic match was defined as the transformation associated with the global minimum (found by an exhaustive search). Without preprocessing, the registration reliability was low (less than 27%). With preprocessing, about 90% of the matches were successful, with a difference with our gold standard (manual registration) of about 1 mm and 1 degree SD. All tested cost functions performed similarly. However, the number of local minima using mutual information was larger than for the other tested cost functions. A cost function based on the mean product of the corresponding pixel values had the least number of local minima. In conclusion, gray value based registration of portal images is applicable for a wide range of treatment sites. However, pre-processing of the images is essential.     

Automatic extraction of corresponding points for the registration of medical images.

In this paper we address the problem of finding corresponding points in a reference and its subsequent image with the aim of registering the images. A whole-image-content-based automatic algorithm for extracting point pairs from 2-D monomodal medical images has been developed. The properties of point distinctiveness, point pair similarity, and point pair consistency have been incorporated into the steps which lead to the automatic extraction and weighting of point pairs. The selection of the most distinctive points of the reference image, and the search for their corresponding points in the subsequent image, have two things in common. First, the local operator by which the distinctive points are selected mimics the template matching used to find the corresponding points. Second, the same similarity measure is used for both tasks. We have applied the algorithm to a variety of computer-generated and real medical images, and have both qualitatively and quantitatively evaluated its performance. The results show that the proposed automatic algorithm for point extraction is accurate and robust and that it may significantly improve on the accuracy, reproducibility, and speed of the manual extraction of corresponding points.     

Evaluation of three types of reference image data for external beam radiotherapy target localization using digital tomosynthesis (DTS)

Digital tomosynthesis (DTS) is a fast, low-dose three-dimensional (3D) imaging approach which yields slice images with excellent in-plane resolution, though low plane-to-plane resolution. A stack of DTS slices can be reconstructed from a single limited-angle scan, with typical scan angles ranging from 10 degrees to 40 degrees and acquisition times of less than 10 s. The resulting DTS slices show soft tissue contrast approaching that of full cone-beam CT. External beam radiotherapy target localization using DTS requires the registration of on-board DTS images with corresponding reference image data. This study evaluates three types of reference volume: original reference CT, exact reference DTS (RDTS), and a more computationally efficient approximate reference DTS (RDTSapprox), as well as three different DTS scan angles (22 degrees, 44 degrees, and 65 degrees) for the DTS target localization task. Three-dimensional mutual information (MI) shared between reference and onboard DTS volumes was computed in a region surrounding the spine of a chest phantom, as translations spanning +/-5 mm and rotations spanning +/-5 degrees were simulated along each dimension in the reference volumes. The locations of the MI maxima were used as surrogates for registration accuracy, and the width of the MI peaks were used to characterize the registration robustness. The results show that conventional treatment planning CT volumes are inadequate reference volumes for direct registration with on-board DTS data. The efficient RDTSapprox method also appears insufficient for MI-based registration without further refinement of the technique, though it may be suitable for manual registration performed by a human observer. The exact RDTS volumes, on the other hand, delivered a 3D DTS localization accuracy of 0.5 mm and 0.50 along each axis, using only a single 44 degrees coronal on-board DTS scan of the chest phantom.     

Automatic noise robust registration of radiographs for subtraction using strategic local correlation: an application to radiographs of dental implants

Most of digital subtraction methods in dental radiography are based on registration using manual landmarks. We have developed an automatic registration method without using the manual selection of landmarks. By restricting a geometrical matching of images to a region of interest (ROI), we compare the cross-correlation coefficient only between the ROIs. The affine or perspective transform parameters satisfying maximum of cross-correlation between the local regions are searched iteratively by a fast searching strategy. The parameters are searched on the 14 scale image coarsely and then, the fine registration is performed on the original scale image. The developed method can match the images corrupted by Gaussian noise with the same accuracy for the images without any transform simulation. The registration accuracy of the perspective method shows a 17% improvement over the manual method. The application of the developed method to radiographs of dental implants provides an automatic noise robust registration with high accuracy in almost real time.     

Automatic lung nodule matching on sequential CT images

We propose an automatic segmentation and registration method that provides more efficient and robust matching of lung nodules in sequential chest computed tomography (CT) images. Our method consists of four steps. First, the lungs are extracted from chest CT images by the automatic segmentation method. Second, gross translational mismatch is corrected by optimal cube registration. This initial alignment does not require extracting any anatomical landmarks. Third, the initial alignment is step-by-step refined by hierarchical surface registration. To evaluate the distance measures between lung boundary points, a three-dimensional distance map is generated by narrow-band distance propagation, which drives fast and robust convergence to the optimal value. Finally, correspondences of manually detected nodules are established from the pairs with the smallest Euclidean distances. Experimental results show that our segmentation method accurately extracts lung boundaries and the registration method effectively finds the nodule correspondences.     

基于自适应标记分水岭算法的肝脏CT图像自动分割

目的为减少人工交互提出了基于自适应标记分水岭的CT系列图像肝脏区域自动分割算法。方法首先对图像进行形态学重构运算以平滑图像,然后计算多尺度形态学梯度,同时提出利用梯度图像非零的局部极小值点的均值进行自适应标记提取,以避免分水岭的过分割和欠分割,再结合肝脏为最大的实质性脏器和相邻图像的相似性实现CT系列图像的肝区自动分割。结果该算法能自动、快速地提取CT系列图像中的肝脏区域。结论分水岭算法能准确定位区域的边缘,通过选择合适的阈值对梯度图像进行标记以抑制分水岭的过分割,实现医学图像中感兴趣区域的自动分割。     

Automated cerebrum segmentation from three-dimensional sagittal brain MR images

We present a fully automated cerebrum segmentation algorithm for full three-dimensional sagittal brain MR images. First, cerebrum segmentation from a midsagittal brain MR image is performed utilizing landmarks, anatomical information, and a connectivity-based threshold segmentation algorithm as previously reported. Recognizing that cerebrum in laterally adjacent slices tends to have similar size and shape, we use the cerebrum segmentation result from the midsagittal brain MR image as a mask to guide cerebrum segmentation in adjacent lateral slices in an iterative fashion. This masking operation yields a masked image (preliminary cerebrum segmentation) for the next lateral slice, which may truncate brain region(s). Truncated regions are restored by first finding end points of their boundaries, by comparing the mask image and masked image boundaries, and then applying a connectivity-based algorithm. The resulting final extracted cerebrum image for this slice is then used as a mask for the next lateral slice. The algorithm yielded satisfactory fully automated cerebrum segmentations in three-dimensional sagittal brain MR images, and had performance superior to conventional edge detection algorithms for segmentation of cerebrum from 3D sagittal brain MR images.     

基于图像域的PROPELLER磁共振仿射运动校正新算法

在对人体腹部或其他软组织进行PROPELLER磁共振成像(MRI)过程中,成像结果会有不同程度的仿射运动伪影。迄今为止提出的比较好的仿射运动校正方法是基于频率域的,即用仿射运动信息对k空间条进行校正后网格化重建得到最终图像。但结果中会有混叠现象,并且在一些细节上有伪影存在。本文提出一种新的PRO-PELLER仿射校正方法,将仿射运动模型加入到基于图像域的PROPELLER磁共振图像重建过程中,首先在图像域通过图像配准算法获得仿射运动信息,然后利用仿射信息校正k空间坐标从而完成对k空间条的密度补偿,再通过逆傅立叶变换得到各个子图像并在图像域进行仿射校正,最后通过旋转完成线性叠加得到最终结果。仿真实验表明,相对于现有的PROPELLER仿射校正算法,本文所提算法对于仿射运动造成的伪影具有更好的校正效果。     

颅面骨三维定量测量及其临床意义

目的：研究当下颌骨处于正常牙尖交错位时,正常成人颅面骨三维空间的相对位置关系。方法：选择６０名正常范围的志愿者,在正常牙尖交错位状态下行螺旋ＣＴ扫描,三维影像重建。在ＥＡＳＹＶＩＳＩＯＮ工作站上对重建的颅面骨影像进行三维测量。在三维骨结构上定点３７个,对有意义的两点间直线距离三点所形成的角度,计算机自动测量,并显示数据。结果：颅面骨ＣＴ三维测量较为客观和精确地反映出在正常牙尖交错位时颅面骨三维空间的相对位置关系,数据比值的标准差远小于数据范围的标准差。结论：计算机三维定量测量精确性高。本研究提供正常数据,为颅骨骨畸形患者提供参考,颅面骨三维重建有助于正颌外科诊断和确定手术方案术前后对比。     

An observer study comparing spot imaging regions selected by radiologists and a computer for an automated stereo spot mammography technique

We are developing an automated stereo spot mammography technique for improved imaging of suspicious dense regions within digital mammograms. The technique entails the acquisition of a full-field digital mammogram, automated detection of a suspicious dense region within that mammogram by a computer aided detection (CAD) program, and acquisition of a stereo pair of images with automated collimation to the suspicious region. The latter stereo spot image is obtained within seconds of the original full-field mammogram, without releasing the compression paddle. The spot image is viewed on a stereo video display. A critical element of this technique is the automated detection of suspicious regions for spot imaging. We performed an observer study to compare the suspicious regions selected by radiologists with those selected by a CAD program developed at the University of Michigan. True regions of interest (TROIs) were separately determined by one of the radiologists who reviewed the original mammograms, biopsy images, and histology results. We compared the radiologist and computer-selected regions of interest (ROIs) to the TROIs. Both the radiologists and the computer were allowed to select up to 3 regions in each of 200 images (mixture of 100 CC and 100 MLO views). We computed overlap indices (the overlap index is defined as the ratio of the area of intersection to the area of interest) to quantify the agreement between the selected regions in each image. The averages of the largest overlap indices per image for the 5 radiologist-to-computer comparisons were directly related to the average number of regions per image traced by the radiologists (about 50% for 1 region/image, 84% for 2 regions/image and 96% for 3 regions/image). The average of the overlap indices with all of the TROIs was 73% for CAD and 76.8% +/- 10.0% for the radiologists. This study indicates that the CAD determined ROIs could potentially be useful for a screening technique that includes stereo spot mammography imaging.     

A physical model of multiple-image radiography

We recently proposed a phase-sensitive x-ray imaging method called multiple-image radiography (MIR), which is an improvement on the diffraction-enhanced imaging technique. MIR simultaneously produces three images, depicting separately the effects of absorption, refraction and ultra-small-angle scattering of x-rays, and all three MIR images are virtually immune to degradation caused by scattering at higher angles. Although good results have been obtained using MIR, no quantitative model of the imaging process has yet been developed. In this paper, we present a theoretical prediction of the MIR image values in terms of fundamental physical properties of the object being imaged. We use radiative transport theory to model the beam propagation, and we model the object as a stratified medium containing discrete scattering particles. An important finding of our analysis is that the image values in all three MIR images are line integrals of various object parameters, which is an essential property for computed tomography to be achieved with conventional reconstruction methods. Our analysis also shows that MIR truly separates the effects of absorption, refraction and ultra-small-angle scattering for the case considered. We validate our analytical model using real and simulated imaging data.     

Fast and Accurate Registration Techniques for Affine and Nonrigid Alignment of MR Brain Images

Registration of magnetic resonance brain images is a geometric operation that determines point-wise correspondences between two brains. It remains a difficult task due to the highly convoluted structure of the brain. This paper presents novel methods, Brain Image Registration Tools (BIRT), that can rapidly and accurately register brain images by utilizing the brain structure information estimated from image derivatives. Source and target image spaces are related by affine transformation and non-rigid deformation. The deformation field is modeled by a set of Wendland’s radial basis functions hierarchically deployed near the salient brain structures. In general, nonlinear optimization is heavily engaged in the parameter estimation for affine/non-rigid transformation and good initial estimates are thus essential to registration performance. In this work, the affine registration is initialized by a rigid transformation, which can robustly estimate the orientation and position differences of brain images. The parameters of the affine/non-rigid transformation are then hierarchically estimated in a coarse-to-fine manner by maximizing an image similarity measure, the correlation ratio, between the involved images. T1-weighted brain magnetic resonance images were utilized for performance evaluation. Our experimental results using four 3-D image sets demonstrated that BIRT can efficiently align images with high accuracy compared to several other algorithms, and thus is adequate to the applications which apply registration process intensively. Moreover, a voxel-based morphometric study quantitatively indicated that accurate registration can improve both the sensitivity and specificity of the statistical inference results.     

Quantitative evaluation of free-form deformation registration for dynamic contrast-enhanced MR mammography

In this paper, we present an evaluation study of a set of registration strategies for the alignment of sequences of 3D dynamic contrast-enhanced magnetic resonance breast images. The accuracy of the optimal registration strategies was determined on unseen data. The evaluation is based on the simulation of physically plausible breast deformations using finite element methods and on contrast-enhanced image pairs without visually detectable motion artifacts. The configuration of the finite element model was chosen according to its ability to predict in vivo breast deformations for two volunteers. We computed transformations for ten patients with 12 simulated deformations each. These deformations were applied to the postcontrast image to model patient motion occurring between pre- and postcontrast image acquisition. The original precontrast images were registered to the corresponding deformed postcontrast images. The performance of several registration configurations (rigid, affine, B-spline based nonrigid, single-resolution, multi-resolution, and volume-preserving) was optimized for five of the ten patients. The images were most accurately aligned with volume-preserving single-resolution nonrigid registration employing 40 or 20 mm control point spacing. When tested on the remaining five patients the optimal configurations reduced the average mean registration error from 1.40 to 0.45 mm for the whole breast tissue and from 1.20 to 0.32 mm for the enhancing lesion. These results were obtained on average within 26 (81) min for 40 (20) mm control point spacing. The visual appearance of the difference images from 30 patients was significantly improved after 20 mm volume-preserving single-resolution nonrigid registration in comparison to no registration or rigid registration. No substantial volume changes within the region of the enhancing lesions were introduced by this nonrigid registration.     

基于特征点的人体红外图像自动分割技术

目的医学红外人体图像区域分割是大规模医学红外图像处理的关键步骤。为快速有效地获取医学红外图像中的人体信息,本文提出一种在医学红外图像中自动提取并划分人体区域的方法。方法由红外热像仪在静室中采集人的裸体红外图像,然后通过对红外人体图像灰度分布特征分析而取得的阈值来获取人体区域,以人体横向距离(宽度)函数结合人体红外图像中的特殊方向亮带的识别,提取人体的特征点,并通过特征点对人体区域进行分割。结果对来自8人的72幅图像进行验证,其中64幅可以正确分割,证明该方法可以对直立姿势的红外人体图像进行自动区域分割与提取。结论该红外人体图像区域自动分割算法可为基于红外图像的疾病筛查及计算机辅助诊断提供技术基础。     

Application of Kohonen network for automatic point correspondence in 2D medical images

In this paper, a generalized application of Kohonen Network for automatic point correspondence of unimodal medical images is presented. Given a pair of two-dimensional medical images of the same anatomical region and a set of interest points in one of the images, the algorithm detects effectively the set of corresponding points in the second image, by exploiting the properties of the Kohonen self organizing maps (SOMs) and embedding them in a stochastic optimization framework. The correspondences are established by determining the parameters of local transformations that map the interest points of the first image to their corresponding points in the second image. The parameters of each transformation are computed in an iterative way, using a modification of the competitive learning, as implemented by SOMs. The proposed algorithm was tested on medical imaging data from three different modalities (CT, MR and red-free retinal images) subject to known and unknown transformations. The quantitative results in all cases exhibited sub-pixel accuracy. The algorithm also proved to work efficiently in the case of noise corrupted data. Finally, in comparison to a previously published algorithm that was also based on SOMs, as well as two widely used techniques for detection of point correspondences (template matching and iterative closest point), the proposed algorithm exhibits an improved performance in terms of accuracy and robustness.     

Medical Image Tamper Detection Based on Passive Image Authentication

Telemedicine has gained popularity in recent years. Medical images can be transferred over the Internet to enable the telediagnosis between medical staffs and to make the patient’s history accessible to medical staff from anywhere. Therefore, integrity protection of the medical image is a serious concern due to the broadcast nature of the Internet. Some watermarking techniques are proposed to control the integrity of medical images. However, they require embedding of extra information (watermark) into image before transmission. It decreases visual quality of the medical image and can cause false diagnosis. The proposed method uses passive image authentication mechanism to detect the tampered regions on medical images. Structural texture information is obtained from the medical image by using local binary pattern rotation invariant (LBPROT) to make the keypoint extraction techniques more successful. Keypoints on the texture image are obtained with scale invariant feature transform (SIFT). Tampered regions are detected by the method by matching the keypoints. The method improves the keypoint-based passive image authentication mechanism (they do not detect tampering when the smooth region is used for covering an object) by using LBPROT before keypoint extraction because smooth regions also have texture information. Experimental results show that the method detects tampered regions on the medical images even if the forged image has undergone some attacks (Gaussian blurring/additive white Gaussian noise) or the forged regions are scaled/rotated before pasting.     

Temporal subtraction in chest radiography: automated assessment of registration accuracy

Radiologists routinely compare multiple chest radiographs acquired from the same patient over time to more completely understand changes in anatomy and pathology. While such comparisons are achieved conventionally through a side-by-side display of images, image registration techniques have been developed to combine information from two separate radiographic images through construction of a "temporal subtraction image." Although temporal subtraction images provide a powerful mechanism for the enhanced visualization of subtle change, errors in the clinical evaluation of these images may arise from misregistration artifacts that can mimic or obscure pathologic change. We have developed a computerized method for the automated assessment of registration accuracy as demonstrated in temporal subtraction images created from radiographic chest image pairs. The registration accuracy of 150 temporal subtraction images constructed from the computed radiography images of 72 patients was rated manually using a five-point scale ranging from "5-excellent" to "1-poor;" ratings of 3, 4, or 5 reflected clinically acceptable subtraction images, and ratings of 1 or 2 reflected clinically unacceptable images. Gray-level histogram-based features and texture measures are computed at multiple spatial scales within a "lung mask" region that encompasses both lungs in the temporal subtraction images. A subset of these features is merged through a linear discriminant classifier. With a leave-one-out-by-patient training/testing paradigm, the automated method attained an A(z) value of 0.92 in distinguishing between temporal subtraction images that demonstrated clinically acceptable and clinically unacceptable registration accuracy. A second linear discriminant classifier yielded an A(z) value of 0.82 based on a feature subset selected from an independent database of digitized film images. These methods are expected to advance the clinical utility of temporal subtraction images for chest radiography.