Multiple Kernel Learning SVM And Lung Nodule Recognition

Designing suitable kernel functions for a given problem is the core issue ofsupport vector machine (SVM) and kernel learning methods. Because a single kernelfunction has a fixed format and its changing space is relatively small, thegeneralization capability and robustness of single kernel function SVMs are limited.Current SVM algorithms mostly use single kernel functions, which are notapplicable to all the specific problems. Compared with single kernel functions,multiple kernel learning methods can overcome the problems such as heterogeneousinformation in the sample characteristics, huge sample size, irregularmulti-dimentional data, and unflat data distribution in the high-dimensional featurespace. The key to improving the performance of SVM is to design the parameters ofmultiple kernel function flexibly.Pulmonary nodule recognition is the core module of the computer aideddiagnosis (CAD) for pulmonary nodule detection. The main task of pulmonary nodulerecognition algorithm is to remove the non-nodular areas in the lung nodule candidateregions of interest and try to recognize all lung nodules, which requires the highaccuracy and sensitivity indicators. While SVMs are the main method of thepulmonary nodules identification now, they mostly use single kernel functions, whichmakes it hard to balance multiple detection indicators.The main research purpose of this work is to combine the multiple kernellearning methods with different forms of SVM algorithms, and to validate thecombined method by applying it to lung nodule recognition. The detailed tasks of thework are outlined as follows:1. An SVM learning method based on hybrid kernel functions is explored toidentify the benign and malignant lung nodules. By the5-fold cross-validation andthe optimal parameter test set based on the best accuracy indicators, the sensitivity ofthe mixed kernel SVM algorithm is up to92.59%, and the accuracy is up to92%.Compared with the other single kernel SVM methods, the proposed method canbetter combine these two indicators, which has high robustness and strong ability to identify lung nodules.2. A new way of thought is provided which allows the two-dimensional digitalimages as input mode in Matrix Least Squares Support Vector Machines (MatLSSVM)algorithm.40suspected lung nodule regions of interested (ROI), ROI are selected inthe experiments, including20positive samples and20negative samples. In theMatLSSVM algorithm, a linear kernel is used and the parameters are selected by gridsearching method. The optimal parameter is selected by10-fold cross-validation andthe recognition results are obtained: the accuracy is97.5%, the sensitivity is100%and the specificity result is95%. The results show that the identification oftrue-positive nodules is not missed; the detection rate of the true-negative is thehighest. Obviously, the detection works the best.3. Based on MatLSSVM, an Multiple Kernel Learning method based onMatLSSVM (MKL-MatLSSVM) is proposed to solve nonlinear division andtwo-dimensional input mode classification problems. The experimental results of thefinal test on the test set show that, the sensitivity of the algorithm MatLSSVM is up to90%, accuracy is up to93.13%, and the specificity is up to94.17%. The threeindicators are all optimal. When the weight factors and each kernel parameter of theMKL-MatLSSVM algorithms are all set to specific values, the kernel functions ofmost situations are covered, and the MatLSSVM algorithm become a special case ofMKL-MatLSSVM algorithm. Compared with the existing classical algorithms, thisalgorithm has the highest accuracy. In addition, the Receiver Operating CharacteristicCurve (ROC) areas corresponding to the least squares support vector machinealgorithm covering mixed kernel and (Radial Basis Function) RBF kernel matrix arealso the largest. The validity of the algorithm is verified.4. To deal with the situations with imbalanced positive and negative samples，thepenalty for positive sample is increased, wheares the original penalty for negativesamples is retained. On the basis of the above hybrid kernel SVM algorithm, acost-sensitive hybrid kernel SVM algorithm is proposed. Compared with the previoushybrid kernel SVM algorithm, as well as a variety of single-kernel functioncost-sensitive and traditional SVM algorithms, the validity of the cost-sensitive hybridkernel SVM algorithm for imbalanced data set is verified. Effectiveness and betterACC and sensitivity (SEN) indicators are obtained. However, specificity (SPC)indicators are not always optimal in all kernels and algorithms, which implies thatupgrading the SEN indicators comes with downgrading the SPC indicators. This isconsistent with the principles of cost-sensitive SVM algorithms, i.e., obtaining good overall effect by decision boundary translation. In clinical lung nodule recognition,SEN indicators are often paid more attention so as to recognize all nodules. Therefore,the consideration of using cost-sensitive hybrid kernel SVM algorithm in lung nodulerecognition improves SEN indicators, but it may also results in some compromisedSPE indicators.A variety of multiple kernel learning SVM algorithms are compared with theexisting single kernel algorithm in accuracy, sensitivity, specificity and ROC curve.The experimental results demonstrate the reliability and validity of the several kindsof SVM algorithms based on kernel learning proposed in this paper.