Search

Mammogram is currently the best way for early detection of breast cancer. Mass is a typical sign of breast cancer, and the classification of masses as malignant or benign may assist radiologists in reducing the biopsy rate without increasing false negatives. Typically, different geometry and texture features are extracted and utilized to train a classifier to classify a mass. However, not each feature is equally important for a classifier, and some features may indeed decrease the performance of a classifier. In this paper, we investigated the usage of semi-supervised feature selection method for classification. After a mass is extracted from a ROI (region of interest) with level set method. Morphological and texture features are extracted from the segmented regions and surrounding regions. SSLFE (Semi- Supervised Local Feature Extraction, proposed in our previous work) is utilized to select important features for KNN classifier. Mammography images from DDSM were used for experiment. The experimental result shows that by incorporating information embedded in unlabeled data, SSLFE can improve the performance compared to the method without feature selection and traditional Relief method.

Mass classification is an important problem in breast cancer diagnosis. In this paper, we investigated the classification of masses with feature selection. Based on the initial contour guided by radiologist, level set algorithm is used to deform the contour and achieves the final segmentation. Morphological features are extracted from the boundary of segmented regions. Then, important features are extracted based on mutual information criterion. Linear discriminant analysis and support vector machine are investigated for the final classification. Mammography images from DDSM were used for experiment. The method achieved an accuracy of 86.6% with mutual information based feature selection and SVM classifier. The experimental result shows that mutual information based feature selection is useful for the diagnosis of masses.

Mass localization is important in computer-aided detection (CAD) system for the diagnosis of suspicious regions in mammograms. In this paper, we present an automated classification system for the detection of masses in mammographic images. Suspicious regions are located with an adaptive region growing firstly. Then, the initial regions are further refined with narrow band based active contour, which can improve the segmentation accuracy of masses. CLBP (Complete Local Binary Pattern) texture features are extracted from the ROIs (regions of interest) containing the segmented suspicious regions. Finally, the ROIs are classified by means of support vector machine (SVM), with supervision provided by the radiologist's diagnosis. The method was evaluated on a dataset with 231 images, containing 245 masses. Among them, 125 images containing 133 masses are used to optimize the parameters and are used to train SVM. The remaining 106 images are used to test the performance. It obtained 1.36 FPsI at the sensitivity 76.8%. It shows that the proposed method is a promising approach to achieve low FPsI while maintain a high sensitivity.

Classification of mammographic masses as malignant or benign may assist radiologists in reducing the biopsy rate without increasing false negatives. In this paper, we investigated the classification of masses with level set segmentation and multiple kernel learning. Based on the initial contour guided by radiologist, level set segmentation is used to deform the contour and achieves the final segmentation. Morphological features are extracted from the boundary of segmented regions. Linear discriminant analysis, support vector machine and multiple kernel learning are investigated for the final classification. Mammography images from DDSM were used for experiment. The method based on the level set segmentation and the morphological features achieved an accuracy of 76%. The experimental result shows that level set based segmentation can improve the characterization of masses compared with manually rough segmentation.

Dimension reduction is critical in many areas of pattern classification and machine learning and many algorithms have been proposed. Pairwise Covariance-preserving Projection Method (PCPM) was an effective dimension reduction which maximizes the class discrimination and also preserves approximately the pairwise class covariances. A shortcoming of PCPM is that it can only be applied when all labels are given, thus a typical supervised method. Semi-supervised has attracted much attention in recent years since they can utilize both labeled and unlabeled data. In this paper, we extend PCPM to semi-supervised setting. The labeled data points are used to maximize the separability between different classes and the unlabeled data points are used to estimate the intrinsic geometric structure of the data. Specifically, we aim to learn a discriminant function which is as smooth as possible on the data manifold. The target optimization problem involved can be solved efficiently by eigenvalue decomposition. Experimental results on several datasets demonstrate the effectiveness of our method.

Mammography is currently regarded as one of the best ways to detect breast cancer in the early stage. However, due to the limitation in imaging condition and the subtleness of the difference between normal and abnormal features, it is generally difficult to interpret the mammograms. Thus, image enhancement techniques have been widely used in screening mammograms. In this paper, a multiscale contrast enhancement algorithm based on Laplacian Pyramid is developed to enhance the contrast of the mammograms and improve the discernibility of the abnormal features. In the proposed algorithm, an image is first decomposed into a multi-level Laplacian Pyramid and then the enhancement is performed in the reconstruction stage. A multiscale contrast measure is used to modify the coefficients iteratively level by level and the enhanced image is obtained at the lowest level. Experiments proved the effectiveness of the proposed algorithm.

The correct segmentation of textural pattern into different meaningful regions is one of the most important problems in automatic texture image recognition. In this paper, we presented a variational integration of shape prior statistics into a phase-field based segmentation process. By derivating the new phase field functionals with gradient shape policy, we obtain the interface evolution process for texture pattern. The interface evolving uses information from shape templates to guide its behavior to desired features. Experimental results on noisy texture images demonstrated the efficacy of the proposed algorithm.

Dimension reduction is very important for applications in data mining and machine learning. Dependence maximization based supervised feature extraction (SDMFE) is an effective dimension reduction method proposed recently. A shortcoming of SDMFE is that it can only use labeled data, and does not work well when labeled data are limited. However, in many applications, it is a common case. In this paper, we propose a novel feature extraction method, called Semi-Supervised Dependence Maximization Feature Extraction (SSDMFE), which can utilize simultaneously both labeled and unlabeled data to perform feature extraction. The labeled data are used to maximize the dependence and the unlabeled data are used as regulations with respect to the intrinsic geometric structure of the data. Experiments on several datasets are presented and the results demonstrate that SSDMFE achieves much higher classification accuracy than SDMFE when the amount of labeled data are limited.

Local Feature Extraction (LFE) algorithm is an effective feature extraction method developed in recent years. One of the shortcomings of the current LFE algorithm is that it can only process labeled data, and does not work well when the amount of the labeled data is limited. However, it is usually easy to obtain large amount of unlabeled data but only a few labeled data. In this paper, we propose a new feature extraction algorithm, called Semi-Supervised LFE (SSLFE), which can handle both labeled and unlabeled data to perform feature extraction. In the proposed algorithm, the labeled data are used to maximize the margin and the unlabeled data are used as regulations with respect to the intrinsic geometric structure of the data. The final projection matrix can be obtained by eigenvalue decomposition. Experiments on several datasets demonstrate that SSLFE achieves much higher classification accuracy than LFE.

Dimensionality reduction is among the keys in many fields, most of the traditional method can be categorized as local or global ones. In this paper, we consider the dimension reduction problem with prior information is available, namely, semi-supervised dimension reduction. A new dimension reduction method that can explore both the labeled and unlabeled information in the dataset is proposed. The labeled data points are used to maximize the separability between different classes and the unlabeled data points are used to estimate the intrinsic geometric structure of the data. Specifically, we aim to learn a discriminant function which is as smooth as possible on the data manifold. The target optimization problem involved can be solved efficiently with eigenvalue decomposition. Experimental results on several datasets demonstrate the effectiveness of our method.

Locality sensitive discriminant analysis (LSDA) algorithm is a new data analysis tool for studying the class relationship between data points, which can utilize local geometry structure of the data manifold and discriminant information at the same time. A major disadvantage of LSDA is it that can only deal with vector data, and thus is often confronted with singularity problem. In this paper, an extension of LSDA is proposed, called two-dimensional locality sensitive discriminant analysis (2DLSDA), which is directly based on 2D image matrices for face recognition, can overcome the singularity problem and utilize the spatial information among pixels more effectively. Besides, based on the Schur decomposition, the projection matrices can be obtained efficiently with high numerical stability, and orthogonality of projection matrix is guaranteed. Experiments on both ORL and Yale datasets demonstrate that the proposed method can achieve better performance than PCA, LDA and LSDA methods.

Dimension reduction is critical in many areas of pattern classification and machine learning and many discriminant analysis algorithms have been proposed. In this paper, a Pairwise Covariance-preserving Projection Method (PCPM) is proposed for dimension reduction. PCPM maximizes the class discrimination and also preserves approximately the pairwise class covariances. The optimization involved in PCPM can be solved directly by eigenvalues decomposition. Our theoretical and empirical analysis reveals the relationship between PCPM and Linear Discriminant Analysis (LDA), Sliced Average Variance Estimator (SAVE), Heteroscedastic Discriminant Analysis (HDA) and Covariance preserving Projection Method (CPM). PCPM can utilize class mean and class covariance information at the same time. Furthermore, pairwise weight scheme can be incorporated naturally with the pairwise summarization form. The proposed methods are evaluated by both synthetic and real-world datasets.

LTSA (local tangent space alignment) is a recently proposed method for manifold learning, which can efficiently learn nonlinear embedding low-dimensional coordinates of high-dimensional data, and can also reconstruct high dimensional coordinates from embedding coordinates. But it ignores the label information conveyed by data samples, and can not be used for classification directly. In this paper, a transductive manifold classification method, called QLAT (LDA/QR and LTSA based Transductive classifier) is presented, which is based on LTSA and TCM-KNN (transduction confidence machine-k nearest neighbor). In the algorithm, local low-dimensional coordinates is constructed using 2-stage LDA/QR method, which not only utilize the label information of sample data, but also conquer the singularity problem of traditional LDA, then the global low-dimensional embedding manifold is obtained by local affine transforms, finally TCM-KNN method is used for classification on the low-dimensional manifold. Experiments on labeled and unlabeled mixed data set illustrate the effectiveness of the method.

This paper presents an adaptive-weighted hidden Markov models (AWHMM) method for classification of 3D models into a set of pre-determinated model classes. Two new features are proposed to capture model surface orientation information. In the method, each model class is represented by four HMMs corresponding to four type features. During the training process, for each type of feature, the feature statistics of each model class and the spatial dynamics are learned by an HMM. During the classification process, characteristics of the test model are analyzed by the HMMs corresponding to each model class. The likelihood scores provided by the HMMs are calculated, and the highest weighted sum score provides the class identification of the test model. Furthermore, with unsupervised learning, each HMM and type weight are adapted with test models, which results in better modeling over time. Based on experiments, the proposed algorithm achieves much better performance than a baseline method and better performance than HMMs using only two features with fixed weights method.

Phase field models provide a well-established framework for the mathematical description of free boundary problems for image segmentation. In phase field models interfaces represent edges of jacquard images and the determination of the edges of jacquard images is the main goal of image segmentation. In this paper, the phase field model was applied to segment and recognize pattern structures of jacquard images. The segmentation was performed in two major steps. Firstly, a pattern extraction and representation was performed by an adaptive mesh generation scheme. For the conjugate gradient method has been successfully used in solving the symmetric and positive definite systems obtained by the finite element approximation of energy functionals, a novel conjugate gradient algorithm was adapted to the minimization of energy functional of discrete phase model. Experimental results show efficiency of our approach.