الفهرس 
Literature ReviewOnly 14 pages are availabe for public view
 |  | from | 111 |  |  |
| | | from | 111 | | |
Abstract
Melanoma is a sort of skin disease that represents more than seventy-five percent of all skin diseases connected to all fatalities. Nonetheless, doctors have demonstrated that the probability rate of patients improves radically with early analysis and diagnosis. This motivated researchers to seek automated techniques that facilitate early diagnosis of skin cancer. Skin lesion segmentation is a significant advance in the analysis and the resulting treatment of melanoma. Automatic lesion segmentation is of major interest for early detection and treatment of skin cancer, because it provides better accuracy and speed, compared to manual analysis. Lately, deep neural networks have provided better results for medical image segmentation, compared to classical approaches based on machine learning.
In this thesis, first an extensive a review of existing deep network architectures that have been suggested to segment skin lesions, pre-processing and post-processing methods with the available datasets that can be used for research in this area, also presented a comparison between the results of different methods used for skin lesion segmentation showing the strengths and weaknesses of each method.
In this thesis, checked the applicability of deep learning approaches: U-NET, Deep RESU-NET, VGG16UNET, U-Net DENSENET121, U-Net EfficientNet-B0, Deeplabv3plus, Inception-ResNet-v2-unet, mobilenetv2_unet, Resnet50_unet, and vgg19_unet to the segmentation of skin lesions to detect lesion boundaries by developing deep learning models that have never been used for skin lesions and producing comparative results that will help detect melanoma and successfully define the lesion boundaries. The architectures have been trained on four different datasets. The architectures were trained on the original data set, and then the four datasets were pre-processed to be used for training the architectures. Utilized different kinds of original image transformations, such as center crop, random rotation 90, grid distortion, horizontal flip, and vertical flip which resulted in an increasing number of images corresponding to the training set. Utilized different standard evaluation metrics such as Jaccard Index, accuracy, recall, precision, and F-measure to evaluate the acquired results. The results show that out of these architectures, the U-Net DENSENET121 architecture outperforms with segmentation Accuracy as high as 97.23% and F1 as 95.96% and Jaccard as 92.42% and Recall as 96.33% in the PH2 dataset. In general, the architecture of U-Net DENSENET121 showed a significant superiority over the rest of the architectures. Even when comparing the architectures that we used in this study with the rest of the other methods, we found that U-Net DENSENET121 architecture gave higher results than the rest of the architectures.
This thesis is a comprehensive study that tries to evaluate the segmentation step in skin lesions using deep learning approaches.