الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 122 |  |  |
| | | from | 122 | | |
Abstract
The use of computer-aided systems to diagnose skin lesions has lately become a common research area. Researchers have recently shown a growing interest in building computer-assisted diagnosis methods. One of the most fatal diseases is skin cancer, particularly melanoma. The great similarity between different skin lesions like melanoma and nevus in skin color images increases the challenge of detection and diagnosis. A reliable automated system for skin lesion classification is required for early detection, to save effort, time, and human life. Dermoscopy is a high-resolution technique that assists doctors in a greater accurate skin cancer diagnosis. Melanoma is a fast-growing, aggressive type of skin cancer. Due to this fact, malignant melanoma remains one of the fastest-growing cancers worldwide. After it metastasizes from its origin into other tissues, the response rate to treatment declines to as low as 5%, and its 10-year survival rate is only about 10%. After it metastasizes, there is no surgical removal option available for treatment. Thus, early detection of malignant melanoma is critically important. Among many types of skin cancer, melanoma has the highest false-negative ratio. Therefore, this thesis proposes three methods for the early detection of malignant melanoma: First, we use image preprocessing techniques and methods to ensure that we get a more accurate result. We use image enhancement, rescaling, hair removal, cropping, and segmentation. The PCA methodology is used to develop an algorithm based on machine learning. The second method is a comparison between two deep learning models ResNet50 and VGG16. And finally, we introduce an approach to merging two pre-trained neural network models. This thesis attempts to achieve the most accurate model to classify and detect skin cancer types from seven different classes using deep learning techniques, ResNet-50, VGG-16, and the merged model of these two techniques through the ”concatenate” function. The performance of the proposed model was evaluated through a set of experiments on the ham10000 dataset. We compared our experiment’s results with previous research and found that our model had higher performance compared to theirs. The proposed system has succeeded in achieving a classification accuracy of 94.1%.