During recent decades, the incident of malignant melanoma as the lethal form of skin cancer has been raised. The occurrence in Australia is much higher than US, UK, and Canada with the cases more than 10,000 diagnosis and annual mortality of 1250 people. The persistent raise of this cancer in the worldwide, the high medical cost and death rate have prioritized the early diagnosis of this cancer. The anticipation and cure of melanoma is strictly relevant to its thickness, if it can be detected early, the survival rate would be increased. Although lots of effort has been made to advance the detection of skin cancers, the challenging concerns still about it.
The computer-based detection systems can improve the diagnosis rate of melanoma by 5–30% in comparison with the naked-eye. Since the visual perception often involve some faults, the necessity of second opinion with higher accuracy and reliability is highlighted. On the other hand, it reduces the task and responsibilities that are performed by physicians.
Many researches have been developed in automated detection of melanoma. The potential advantages of such studies are significant and incalculable. Moreover, the difficulties entangle are a lot, and the new contributions in the area are highly appreciated. However, it is extensively acknowledged that the more trustful and reliable detection systems require higher accuracy. The purpose of this thesis is to propose an algorithm for skin cancer diagnosis that is able to classify lesions as malignant or benign automatically.
The different components in an automated diagnosis of skin cancer includes: Pre-processing, segmentation, feature extraction and selection, and classification. In this thesis, after selecting the best image enhancement techniques which are achieved by applying and comparing different noise removal and contrast enhancement techniques on images, the segmentation stage is performed. In this stage, a fully automated segmentation algorithm in dermoscopy images based on k-mean and level-set algorithms are proposed and compared with other algorithms mentioned in this thesis using statistical tools. Proposed algorithm shows the improvement in the results.
In the next stage, after extracting the various features of images, a fully automated feature selection algorithm, Smart PSO-SVM, which optimizes the feature selection stage, is proposed. Comparative study of proposed algorithm with other algorithms is performed to analyse the performance of proposed algorithm among others. The results obtained in the best subset of features which feed the classification stage. In classification stage, the use of SA-SVM as a new classifier in the area of skin cancer detection systems is proposed. The average accuracy and F-score are estimated as 87.0611% and 0.9167 respectively. The statistical evaluation using t-test also shows the superiority of proposed algorithm when compares with other algorithms in this thesis.