Registration and difference analysis of corresponding mammogram images

The paper presents Echo State Network (ESN) as classifier to diagnose the abnormalities in mammogram images. Abnormalities in mammograms can be of different types. An efficient system which can handle these abnormalities and draw correct diagnosis is vital. We experimented with wavelet and Local Energy based Shape Histogram (LESH) features combined with Echo State Network classifier. The suggested system produces high classification accuracy of 98% as well as high sensitivity and specificity rates. We compared the performance of ESN with Support Vector Machine (SVM) and other classifiers and results generated indicate that ESN can compete with benchmark classifier and in some cases beat them. The high rate of Sensitivity and Specificity also signifies the power of ESN classifier to detect positive and negative case correctly.

In this paper, a concise method is proposed to locate the nipple position on a mammogram image. Mammogram registration is an important preprocessing technique, which can help in finding asymmetrical regions in left or the corresponding right breast. In particular, correct nipple position is the crucial key point of mammogram registration since it is still the most consistent and stable landmark on a mammogram. This work first presents an algorithm of maximum height of the breast border (MHBB) and then proposes two novel approaches, local spatial-maximum mean intensity (MMI) and local maximum zero-crossing (MZC) based on the second-order derivative, finally a combined process depending on the MMI and MZC is obtained. The 213 mammogram images from MIAS and 200 ones from DDSM database are tested for estimating the proposed method. Consequently, the mean Euclidean distance (MED) between the nipple position detected and the ground truth identified by the radiologist is 0.63 cm, within 1 cm of the gold standard. The experimental results hence indicate that our combined process can detect the nipple positions more accurately, as compared to other previous methods. Moreover, the proposed LVNM (Locate visible-nipple mammograms) algorithm designed with the generalization ability for automatic nipple clustering in the MIAS database has also yielded an identification rate of 99.53%.

In this chapter, computer aided detection and diagnosis for Breast tumor detection and classification in Mammography is proposed. The proposed system based on four phases. In the first phase, mammogram image is segmented to sub images with size 64 × 64 pixel then high intensity value of pixel is defined in this sub image and specifies this intensity as a seed point to region growing (RG) algorithm which used to specify the ROI. In the second phase, texture features were extracted using gray-level co-occurrence matrix (GLCM) and combined with shape features to characterize region of interest (ROI) to normal, benign or malignant. In the third phase, malignant ROIs are diagnosed and specified to aided doctor for decision taking. Finally, different methods for evaluating classifier are used using confusion matrix, kappa coefficient and response receiver operating characteristic curve (ROC). The effectiveness of the proposed system was measured using 322 mammogram images from the mammographic image analysis society (MIAS) database. From experimental results show that, the accuracy obtained from ROC curve analysis is AUC 94 % with standard error 0.11. The experimental results shows that the proposed system can accurately segment the breast region in a large range of digitized mammograms.

We present an approach to detecting masses in temporal and bilateral (left-right) pairs of mammograms. Masses typically appear in a mammogram as well-defined, relatively bright regions. The converse is not true; by no means all relatively bright, circumscribed regions correspond to masses. Common to our approach to matching temporal and bilateral mammogram pairs is a representation of the nested structure of “salient” bright regions in a mammogram. The representation is described in Sections 2 and 3. In the case of temporal pairs, “salient” regions are extracted independently in two mammograms of the same breast, nominally the same view at two different times. Those that either appear in the later mammogram but not in the earlier one, or which have changed significantly between the two mammograms we seek to highlight. Similarly, in the case of matching bilateral pairs, salient regions that appear in only one of same-view left and right breast mammograms of the same patient, taken at approximately the same time we seek to highlight. A fuller version of this article appears as Chapter 13 of [4].

Breast cancer is a common and devastating form of cancer, with an estimated 22,700 new cases in 2009 in Canada alone. X‐ray mammography is the most commonly used imaging technique for detection and diagnosis, while magnetic resonance imaging(MRI) is used in some challenging cases. Both modalities rely on different properties of the tissue to form images, and thus contribute different and complimentary information about the breast. However, due to geometric distortions during the acquisition processes, it is difficult to identify and compare the same anatomical location on both modalities. In this work, a method to register 2D mammograms to projection images of MRI volumes is presented. In order to compare the 3D MRI to the 2D mammogram, a “simulated mammogram” is formed from the MRI volume. Three anatomical landmarks on the surface of the breast are identified on each image and aligned to distort the general shape of the mammogram to match that of the MR projection image. Final registration is then achieved by iteratively applying a non‐linear transformation to the mammogram until the mutual information of the two images is maximized. The registration method was tested on eight pairs of images from two volunteers (two mammographic views from each breast). Results from this small dataset are promising, with an average alignment error of 3.9%, measured as the difference in area between the two registered images. Future work will examine a larger dataset, including pathological cases, and quantification of internal alignment errors.

An adjunctive ultrasound mammography system and associated methods are described in which an ultrasound image being displayed near an x-ray mammogram image may be superimposed thereon or thereunder by a user for facilitating rapid comprehension of breast structures and detection of abnormalities therein. In one preferred embodiment, the x-ray mammogram image corresponds to a standard x-ray mammogram view, and the ultrasound image is a thick-slice image representing a thick-slice or slab-like portion of the breast volume substantially parallel to that standard x-ray mammogram view. In another preferred embodiment, the user is permitted to manually manipulate the registration of the ultrasound image with the x-ray mammogram image. It has been found that the manual registration process, which involves manual vernier adjustments responsive to perceived registration differences, can rapidly increase the radiologist's perception of the breast structures being displayed by both component images. Even though ultrasound images tend to have substantially different textures and feature emphases than x-ray images, the bimodal thick-slice/x-ray image, alone or in conjunction with the vernier registration process, can often expose or clarify tissue structures hidden in the separate component images, and can often obviate or explain certain noticed structures in the component images.

Publication in the conference proceedings of EUSIPCO, Rhodes, Greece, 1998

In this thesis, we focus on two important medical image processing and analysis problems: multimodal image registration and cerebrovascular segmentation. First, for the purpose of improving the robustness of image registration, we propose two novel registration objective functions and a new technique for better optimization performance. The first novel objective function is a combination of the multi-dimensional mutual information and an angle measure on the maximum distance-gradient (MDG) vector field. The MDG vector field is a new spatial feature field designed for registration tasks. It encodes both local edge information and globally defined spatial information related to the intensity difference, the distance, and the direction of a voxel to a certain object boundary point. This objective function then integrates both the magnitude and the orientation information of the MDG feature into the image registration process. The other novel objective function can be obtained if the statistical joint intensity mappings estimated from the pre-aligned training image pairs are available. This a priori knowledge can then be used as reference to guide the multi-resolution based registration process, by minimizing a new Kullback-Leibler distance (KLD) based similarity measure between the observed joint intensity distribution and the reference joint intensity distribution. To achieve better optimization, which evidently affects the robustness of a registration process, we propose a general technique based on random transformation disturbance to assist optimization procedures to avoid local optima, and hence to achieve robust optimization results. The technique is based on different properties of the local neighborhood of a global optimum and those of local optima. Second, this thesis presents a statistical segmentation technique for the extraction of vasculatures in three-dimensional rotational angiography (3D-RA). This method uses maximum intensity projections (MIP) to improve the accuracy and robustness of threshold estimation using the expectation maximization (EM) algorithm. It is fully automatic and computationally efficient. Finally, encouraging experimental results of the proposed approaches on both synthetic and clinical medical images are presented in the thesis.

Atherosclerotic plaque rupture is responsible for the majority of myocardial infarctions and acute coronary syndromes. Rupture is initiated by mechanical failure of the plaque cap, and thus study of the deformation of the plaque in the artery can elucidate the events that lead to myocardial infarction. Intravascular ultrasound (IVUS) provides high resolution in vitro and in vivo cross-sectional images of blood vessels. To extract the deformation field from sequences of IVUS images, a registration process must be performed to correlate material points between image pairs. The objective of this study was to determine the efficacy of an image registration technique termed Warping to determine strains in plaques and coronary arteries from paired IVUS images representing two different states of deformation. The Warping technique uses pointwise differences in pixel intensities between image pairs to generate a distributed body force that acts to deform a finite element model. The strain distribution estimated by image-based Warping showed excellent agreement with a known forward finite element solution, representing the gold standard, from which the displaced image was created. The Warping technique had a low sensitivity to changes in material parameters or material model and had a low dependency on the noise present in the images. The Warping analysis was also able to produce accurate strain distributions when the constitutive model used for the Warping analysis and the forward analysis was different. The results of this study demonstrate that Warping in conjunction with in vivo IVUS imaging will determine the change in the strain distribution resulting from physiological loading and may be useful as a diagnostic tool for predicting the likelihood of plaque rupture through the determination of the relative stiffness of the plaque constituents.

Remotely sensed images are the main source for a variety of mapping and change-detection applications. Images from different satellites are employed in several of these applications. However, each type of these images has a different resolution and orientation. Hence, they need to be co-registered before any meaningful use. The first step in the registration process is to find conjugate points between the images. This paper presents a modified method of the Scott and Longuet-Higgins approach to find conjugate points between different remotely sensed images. In such an algorithm, initially, corner points are automatically extracted in two images, and for each pair of points, a cost value is computed. The cost of corresponding any two points is computed using two-dimensional transformation models and pixel intensities. The cost values are then used to fill a cost matrix, and its singular value decomposition is used to find corresponding points. The algorithm is tested on three pairs of satellite images with different resolutions and orientations. The results show that the approach presented here succeeded in finding 93% of conjugate points between different pairs of satellite images using only the image coordinates through the eight-parameter transformation model. Moreover, the results show that including the image intensities in the matching procedure does change the results significantly.

Registration techniques are still considered challenging tasks to remote sensing users, especially after enormous increase in the volume of remotely sensed data being acquired by an ever-growing number of earth observation sensors. This surge in use mandates the development of accurate and robust registration procedures that can handle these data with varying geometric and radiometric properties. This paper aims to develop the traditional registration scenarios to reduce discrepancies between registered datasets in two dimensions (2D) space for remote sensing images. This is achieved by designing a computer program written in Visual Basic language following two main stages: The first stage is a traditional registration process by defining a set of control point pairs using manual selection, then comput the parameters of global affine transformation model to match them and resample the images. The second stage included matching process refinement by determining the shift value in control points (CPs) location depending on radiometric similarity measure. Then shift map technique was adjusted to adjust the process using 2nd order polynomial transformation function. This function has chosen after conducting statistical analyses, comparing between the common transformation functions (similarity, affine, projection and 2nd order polynomial). The results showed that the developed approach reduced the root mean square error (RMSE) of registration process and decreasing the discrepancies between registered datasets with 60%, 57% and 48% respectively for each one of the three tested datasets.
 

End-to-end affine registration framework for histopathological images with weak annotations

In this chapter, we explore shape representation, registration, and modeling through implicit functions. To this end, we propose novel techniques for global and local registration of shapes through the alignment of the corresponding distance transforms by defining objective functions that minimize metrics between the implicit representations of shapes.Registration methods in the space of implicit functions like the sum of squares differences (SSD), which can account for primitive transformations (similarity), and more advanced methods like mutual information, which are able to handle more generic parametric transformations, are considered. To address local correspondences we also propose an objective function on the space of implicit representations where the displacement field is represented with a free form deformation that can guarantee one-to-one mapping. In order to address outliers as well as introduce confidence in the registration process, we extend our registration paradigm to estimate uncertainties through the formulation of local registration as a statistical inference problem in the space of implicit functions. Validation of the method through various applications is proposed: (i) parametric shape modeling and segmentation through active shapes for medical image analysis, (ii) variable bandwidth non-parametric shape modeling for recognition, and (iii) object extraction through a level set method. Promising results demonstrate the potentials of implicit shape representations.Key wordsDistance transformsimplicit representationssum of squares differencesmutual informationfree form deformationsgradient descent

Point set registration is one of the challenging tasks in computer vision. One critical step is to find the corresponding relationship between the model point set and the scene point set. Existing registration algorithms primarily utilize the information of global and local shape, yet neglect the credibility of corresponding relation, therefore they may lead to the insufficient estimation of spatial transformation. To tackle this problem, we firstly adopt a relative polar coordinate system, it performs spatial pooling operation and further divides the feature extraction region into sub-areas with different scales. Then, based on the Relative Average Distance (RAD) and the Relative Average Offset Angle (RAOA), we propose multi-granular MRGS descriptor to extract visual structures of the point set. The similarity between the model point set and scene point set is then represented by the Gaussian Mixture Model, where the weights can be dynamically adjusted during the process of registration. Finally, we apply the robust mixture re-weighting to reduce the impact of false corresponding pairs and reinforce the weight of correct matching points. Experimental results on synthetic data and medical image data not only show that our method outperform state-of-the-art methods, but also demonstrate the robustness of our method when the non-grid transformation of point sets suffers from deformations, noises and outliers.

Partial point cloud registration is an essential and fundamental component of generating complete 3D shapes, aiming at converting partial scans into a unified coordinate system. However, existing point cloud registration methods suffer from inadequately rich local feature extraction and feature interaction. In addition, these methods still face challenges in modeling the global contextual information of point clouds sufficiently, which limits the improvement in registration effectiveness, especially for partial-to-partial point cloud registration with high noise. To overcome these issues, this paper proposes an enhanced PointNet and assignable weights transformer network (LFA-Net) for partial point cloud registration. The model achieves coarse-to-fine point cloud registration through three core modules. First, the Sufficient Local Feature Extraction Module (LFM) is constructed to extract various local feature information. Then, the Adequate Feature Aggregation Module (FAM) is designed to integrate the feature information from different point clouds. Finally, the Assignable Weights Transformer Module (ATM) is presented to stimulate the model’s global modeling ability during the registration process, enabling the selection of representative points for optimal point cloud registration. Extensive experiments conducted on ModelNet40 using partially overlapping point clouds illustrate the superior registration performance of LFA-Net compared with other state-of-the-art methods. Moreover, Numerous experiments on synthetic and real-word datasets further indicate that LFA-Net also has significant advantages in registering partial point clouds with noise and unseen categories, which demonstrates its excellent robustness and generalization ability for real-world practical application.