| Περίληψη: | This thesis is dealing with classifiers in Computer Aided Diagnosis in medical imaging. In particular, it focuses on artificial neural networks and feature selection methods.
The specific goals of the thesis are:
1. Search for optimal topology of a feed-forward neural network (FFNN), dealing with four (4) medical imaging classification problems (Cytology, MRI, CT, and Mammography).
2. Study three (3) feature selection (dimensionality reduction) methods including PCA, stepwise analysis and t-test ranking for the FFNN topology defined in the previous step, for the four (4) medical imaging classification problems at hand.
3. Compare performance of the FFNN scheme to KNN, SVM, PNN and LDA classifiers, dealing with the above mentioned four (4) medical imaging classification problems. 10-fold cross validation estimation of generalization performance (generalization error) of the classification schemes was utilized.
4. Statistical significance of the results was validated utilizing ANOVA and Duncan’s test.
To facilitate experimentation, a user-friendly application was developed (Chapter 3) that allows the user to find the best network topology on feature vectors, selected by various pre-processing techniques, and compared with other classifiers.
The results of this are:
1. There is no statistical evidence that the different topology that is tested have any impact on classification performance of FFNN in any of the classification problem that this thesis is dealt off.
2. The stepwise method of dimensionality reduction (feature selection) is statistically significance better method than the other methods, except in the case of one dataset (Cytology) where there are no statistical significant differences. This is because of the inherent ability of stepwise method to select uncorrelated features unlike the other two methods (the datasets that the stepwise featured better performance had many highly correlated features).
3. There is no statistical significant better classifier in most cases, while neuronal classifier exhibits very good behaviour on all cases. For that reason, the selection of classifier does not seem to affect the classification problems at hand. Furthermore, the choice of classifier could be done based on other criteria than the classification performance, such as, the simplicity and plasticity, features that characterize the FFNN.
|
|---|
