Dazhe Zhao

An Optimized Cost-Sensitive SVM for Imbalanced Data Learning

Class imbalance is one of the challenging problems for machine learning in many real-world applications. Cost-sensitive learning has attracted significant attention in recent years to solve the problem, but it is difficult to determine the precise misclassification costs in practice. There are also other factors that influence the performance of the classification including the input feature subset and the intrinsic parameters of the classifier. This paper presents an effective wrapper framework incorporating the evaluation measure (AUC and G-mean) into the objective function of cost sensitive SVM directly to improve the performance of classification by simultaneously optimizing the best pair of feature subset, intrinsic parameters and misclassification cost parameters. Experimental results on various standard benchmark datasets and real-world data with different ratios of imbalance show that the proposed method is effective in comparison with commonly used sampling techniques.

Segmentation of the Left Ventricle Using Distance Regularized Two-Layer Level Set Approach

We propose a novel two-layer level set approach for segmentation of the left ventricle (LV) from cardiac magnetic resonance (CMR) short-axis images. In our method, endocardium and epicardium are represented by two specified level contours of a level set function. Segmentation of the LV is formulated as a problem of optimizing the level set function such that these two level contours best fit the epicardium and endocardium. More importantly, a distance regularization (DR) constraint on the level contours is introduced to preserve smoothly varying distance between them. This DR constraint leads to a desirable interaction between the level contours that contributes to maintain the anatomical geometry of the endocardium and epicardium. The negative influence of intensity inhomogeneities on image segmentation are overcome by using a data term derived from a local intensity clustering property. Our method is quantitatively validated by experiments on the datasets for the MICCAI grand challenge on left ventricular segmentation, which demonstrates the advantages of our method in terms of segmentation accuracy and consistency with anatomical geometry.

Ensemble-based hybrid probabilistic sampling for imbalanced data learning in lung nodule CAD

Classification plays a critical role in false positive reduction (FPR) in lung nodule computer aided detection (CAD). The difficulty of FPR lies in the variation of the appearances of the nodules, and the imbalance distribution between the nodule and non-nodule class. Moreover, the presence of inherent complex structures in data distribution, such as within-class imbalance and high-dimensionality are other critical factors of decreasing classification performance. To solve these challenges, we proposed a hybrid probabilistic sampling combined with diverse random subspace ensemble. Experimental results demonstrate the effectiveness of the proposed method in terms of geometric mean (G-mean) and area under the ROC curve (AUC) compared with commonly used methods.

Pulmonary Nodule Classification with Deep Convolutional Neural Networks on Computed Tomography Images

Computer aided detection (CAD) systems can assist radiologists by offering a second opinion on early diagnosis of lung cancer. Classification and feature representation play critical roles in false-positive reduction (FPR) in lung nodule CAD. We design a deep convolutional neural networks method for nodule classification, which has an advantage of autolearning representation and strong generalization ability. A specified network structure for nodule images is proposed to solve the recognition of three types of nodules, that is, solid, semisolid, and ground glass opacity (GGO). Deep convolutional neural networks are trained by 62,492 regions-of-interest (ROIs) samples including 40,772 nodules and 21,720 nonnodules from the Lung Image Database Consortium (LIDC) database. Experimental results demonstrate the effectiveness of the proposed method in terms of sensitivity and overall accuracy and that it consistently outperforms the competing methods.

A method of pulmonary nodule detection utilizing multiple support vector machines

It has been proven that early detection of pulmonary nodules is an important clinical indication for early-stage lung cancer diagnosis. Recently, support vector machines(SVMs) have been extensively used in pattern recognition. However, the application object for SVMs used for false positives(FPs) reduction when detecting lung nodules is generally based on only axial plane. In this paper, we propose a computerized system aimed at lung nodules detection in Multi-Slice Computed Tomography(MSCT) scans with multiple SVMs; it segments the lung field, extracts three sets of candidates regions with two dimensional(2D) dot-enhancement filter on three slice directions respectively, reduces the FPs with multiple SVMs, and then, integrates the classification results by using pixel analysis and region growing method. The proposed scheme is applied on two lung CT datasets. The experimental results illustrate the efficiency of the proposed method.

A multi-kernel based framework for heterogeneous feature selection and over-sampling for computer-aided detection of pulmonary nodules

Classification plays a critical role in False Positive Reduction (FPR) in lung nodule Computer Aided Detection (CAD). To achieve effective recognition of nodule, many machine learning methods have been proposed. However, multiple heterogeneous feature subsets, high dimensional irrelevant features, as well as imbalanced distribution between the nodule and non-nodule classes typically makes this problem challenging. To solve these challenges, we proposed a multi-kernel based framework for feature selection and imbalanced data learning in Lung nodule CAD, involving multiple kernel learning with a ź 2 , 1 norm regularizer for heterogeneous feature fusion and selection from the feature subset level, a multi-kernel feature selection based on pairwise similarities from the feature level, and a multi-kernel over-sampling for the imbalanced data learning. Experimental results demonstrate the effectiveness of the proposed method in terms of Geometric mean (G-mean) and Area under the ROC curve (AUC), and consistently outperform the competing methods. HighlightsProposed a unified multiple kernel framework to classify potential nodule objects.Regularized multiple kernel with l 2 , 1 źnorm to fuse the heterogeneous feature subsets.Two different feature selection from heterogeneous feature subsets.Over-sampling positive instances in kernel space for imbalanced data.

A PSO-Based Cost-Sensitive Neural Network for Imbalanced Data Classification

Learning from imbalanced data is an important and common problem. Many methods have been proposed to address and attempt to solve the problem, including sampling and cost-sensitive learning. This paper presents an effective wrapper approach incorporating the evaluation measure directly into the objective function of cost-sensitive neural network to improve the performance of classification, by simultaneously optimizing the best pair of feature subset, intrinsic structure parameters and misclassification costs. The optimization is based on Particle Swarm Optimization. Our designed method can be applied on the binary class and multi-class classification. Experimental results on various standard benchmark datasets show that the proposed method is effective in comparison with commonly used sampling techniques.

Segmentation of longitudinal brain MR images using bias correction embedded fuzzy c-means with non-locally spatio-temporal regularization

An automated method is proposed to segment brain tissues in longitudinal MR images.The method has an inherent mechanism to deal with intensity inhomogeneities.Spatio-temporal regularization is used to ensure segmentation consistency.Results are consistent in segmenting an arbitrary number of image series. We propose an automated method for segmentation of brain tissues in longitudinal MR images. In the proposed method, images acquired at each time point are first separately segmented into white matter, gray matter, and cerebrospinal fluid by bias correction embedded fuzzy c-means. Intensities differences are then defined as similarities of each voxel to the cluster centroids. After being normalized in inter-class, the similarities are incorporated into a non-local means de-noising formula to regularize the segmentation in both spatial and temporal dimensions. Non-locally regularization results are used to compute final membership functions for the segmentation. To improve time performance, we accelerate the modified de-noising algorithm using CUDA and obtain a 200 × performance improvement. Quantitative comparison with the state-of-the-art methods on BrainWeb dataset demonstrate advantages of the proposed method in terms of segmentation accuracy and the ability to consistently segment brain tissues in an arbitrary number of longitudinal brain MR image series.

Computer aided detection of lung nodules based on voxel analysis utilizing support vector machines

The detection of pulmonary nodules is proven to be of critical importance in early-stage lung cancer diagnosis. Many computer aided detection (CAD) methods combined with morphological approach and pattern recognition technology to identify lung nodules have been proposed to assist the radiologists to improve sensitivity of diagnosis. We present a computer aided lung nodule detection scheme based on analysis of enhanced voxel in three dimensional (3D) CT image. The method is multi-step, including lung fields segmentation, initial nodule candidates enhancement, enhanced voxel feature extraction, voxels classification with support vector machines (SVMs) and nodule decision rule. Two lung nodule data sets are employed to evaluate the performance of the computerized scheme. The experimental results illustrate the efficiency of the proposed method. We intend to improve the voxel enhancement procedure to increase the performance of the scheme.

Image segmentation and bias correction using local inhomogeneous iNtensity clustering (LINC)

Image segmentation is still an open problem due to the existing of intensity inhomogeneity and noise. To accurately segment images with these biases, a local inhomogeneous intensity clustering (LINC) model is proposed. In LINC, a linear combination of a given set of smooth orthogonal basis functions is used to estimate the bias field. A local clustering criterion function is first defined to cluster the nearly homogeneous intensities in a relatively small neighborhood of each pixel. An energy functional is then defined by integrating the function with respect to the neighborhood center. This energy together with a regularization term and an arc length term are incorporated into a variational level set formulation in which de-nosing is implicitly included due to the implied convolution. Image segmentation and bias correction can be simultaneously achieved by updating variables of the final energy functional iteratively till it is stable or a predetermined iteration number is reached. The proposed model LINC has been extensively tested on both synthetic and real images. Experimental results and comparison with state-of-the-art methods demonstrate the advantages of the proposed model in terms of segmentation accuracy, bias field correction, dealing with noise, and robustness to initialization. HighlightsA level set method is proposed for image segmentation and bias correction.The bias field is simulated by linearly combining a given set of basis functions.The proposed method is able to segment images with intensity inhomogeneity.The proposed method is also able to correct the bias field from the image.Experiments demonstrate that our method is also robust to noise and initialization.