The Study Of SVM-based Medical Image Classification

Currently with the wide application of various image equipment in medicine, medical image based computer aided diagnosis (MIBCAD) has developed rapidly. Computer aided diagnosis (CAD) can improve radiologists' accuracy and help them to identify and classify the medical images quickly and efficiently. One of the most important steps of the MIBCAD is the pattern classification based on feature extraction. Several classification methods such as ANN are based on traditional statistical theory with infinite training samples. But in practice, the number of the sample is limited. Therefore, the traditional methods are inclined to bring many problems like overfitting and local minimum. Vipnik proposed support Vector Machine (SVM) in 1992-1995, and it is derived from Vapnik-Chervouenkis Dimension theory and Structural Risk Minimization (SRM) principle in Statistical Learning Theory (SLT). This idea balances between learning accuracy of special training samples and model's predicted capability by limited sample information. The improvement of SVM has successfully solved the puzzles in many other machine learning methods, such as overfitting, non-linear, disaster of dimensionality, local minimum, and so on. So SVM is concerned as a new popular research aspect after pattern identification and ANN in machine learning field.In this thesis, SVM as a new machine learning method is brought into medical image classification. First, we described the theory basis and mathematic model of SVM, especially emphasized its generalization performance and kernel function principle. Then several improved SVM training arithmetic are proposed, and the attributes of them are analyzed. At last we classify the feature sample data extracted from digitized mammograms. Cross-validation method has been implemented to select the kernel function and the parameters of SVM. The obtained results provide high classification performance. Moreover, in this work the performances of the BP classifier and SVM classifier have been evaluated using the ROC methodology. N-class problem is solved by 1-v-l strategy with SVM, and the result can be considered quite well. In this study, every experimental result has proved that SVM classifier is quite promising and has better performance than the past classifiers. It must improve the development of MIBCAD in the future.