SAHIB, Khouloud2024-07-252024-07-252024https://dspace.univ-annaba.dz//handle/123456789/3600Malignant melanoma, an extremely fatal form of skin cancer, is becoming more common, particularly among individuals with fair skin who are exposed to sunlight. Melanoma may be efficiently treated in its early stages, therefore early detection is crucial for increasing survival chances. Dermatologists play an important role in melanoma detection by examining ermoscopic clinical characteristics such as lesion boundaries and pigment networks, which are essential indications of the illness. However, this approach can be difficult and time-consuming due to differences in lesion size and color, low contrast, and the presence of components such as hair, lubricants, and air bubbles that can interfere with an accurate diagnosis, even for experts.The proposed framework aims to improve the accuracy and efficiency of identifying and classifying skin lesions, particularly melanoma, using deep learning methods. The framework involves several stages: preprocessing, skin lesion segmentation, and lesion classification. In the preprocessing stage, a Gaussian filter is applied to enhance the image quality by reducing artifacts such as fog, hairs, marks, and stains. Also, the region of interest is extracted, and unnecessary portions of the image are removed through automatic cropping.The key component responsible for classification is the Inception-ResNet block, which combines the Inception module with a residual neural network, giving the network the flexibility to learn features at various scales. The suggested models are evaluated on four publicly available datasets: ISIC 2016, ISIC 2017, ISIC 2018, and PH2. W-net and ?-net models achieved high accuracy levels (over 97%) on all databases, outperforming existing methods and offering better sensitivity, specificity, Jaccard index, and dice similarity. In contrast, the Inception ResNet network’s performance for classification varies depending on the dataset and evaluation metrics. The results were then compared to current approaches in the literature, demonstrating that our suggested strategy is accurate, and efficient in both the segmentation and classification of skin lesions.PDFenskin cancer; melanoma; Deep learning, encoder-decoder; segmentation; convolutional neural network; classificationThesis