الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 134 |  |  |
| | | from | 134 | | |
Abstract
Skin cancer is one of the most widespread cancers. It has several classes, such as melanoma, which is one of the major death causes and basal cell carcinoma (BCC) which is the utmost incident form of all cancers. In their early stages, medical experts can hardly discriminate them apart from the benign nevus and the benign keratoses lesion (BKL). This inspired the scope of the present thesis aiming to develop a novel computer-aided diagnosis (CAD) system for multi-class skin lesion classification. An image processing-based CAD system was designed and implemented to efficiently classify the four-pre-mentioned skin lesion classes. Subsequently, a proposed novel texture-based feature set, called the cumulative level-difference mean (CLDM) was extracted. It was calculated at different inter-sample distances and different directions (i.e. horizontal, vertical and both diagonals). Moreover, the Asymmetry, Border irregularity, Color variation and Diameter (ABCD) feature set that was originally used to discriminate between the melanoma and nevus was modified to include more skin lesion types. Accordingly, an integration between the novel CLDM features at inter-sample distance d=10 and the modified-ABCD features was ranked using the eigenvector centrality feature ranking method to produce the final feature vector for further classification. In the present work, different classifiers are employed, namely the support vector machine (SVM) with different kernels and the neural network (NN). The classification performance was then evaluated for several arrangements of the combined feature sets by including different numbers of the highest ranked features with several kernels in the SVM. The experimental results established the superiority of the proposed CAD system using cubic SVM with the highest seven ranked features compared to the use of other feature sets. Furthermore, the proposed system achieved 100% classification accuracy, 100% sensitivity, and 100% specificity with the used dataset, which involved in the present thesis.