Skin Cancer Classification Using Deep Learning Techniques رسالة ماجستير

Abstract

One of the most prevalent types of cancer in the world is skin cancer and Melanoma is widely regarded as the most common and potentially deadly skin cancer. Similar to all other types of cancer, early detection increases the likelihood of successful treatment; fortunately, with the advancement of AI in image processing and the availability of skin cancer historical datasets, it is possible to develop a classification model using deep learning algorithms to facilitate and have a more accurate skin cancer detection, this helps lab teams and simultaneously positively affects the length of the period until the disease is detected, in addition to its cost. By the usage of this system, early detection will save effort, time, and human lives. In this work, the CNN model and various transfer learning-based models, such as; ResNet50, VGG16, VGG19, EfficientNetB7, DenseNet169, Xception, and InceptionV3 are proposed. In addition, averaging ensemble approach is performed on various combinations of the models. This helps to increase the performance of cancer detection. The SIIM-ISIC Melanoma Classification Challenge dataset is used to train and test the proposed model. In this work, data augmentation is applied to the datasets to improve the model training process. We will implement augmentation in two phases; first, applying multiple techniques to the malignant class of images to increase the number of images in this class and solve the imbalance problem. Second, conducting random left and right flipping and up and down flipping augmentation methods on both categories of data to make a diverse set of images from existent images. The proposed method's performance is evaluated using well-known quantitative measures, such as; accuracy, precision, recall, and f1-score. The ensemble modeling of five of the best V models, which are; CNN, EfficientNetB7, VGG16, VGG19, and DenseNet169 outperforms all other models, with an accuracy of 93.79%, a precision of 94.18%, recall of 93.79%, f1 score of 94.77% and AUC score 93.8% on the unseen test dataset. The findings show that the proposed model is more efficient and reliable for automating skin cancer classification in comparison to previous works, shortens the recognition processes, and is valuable in its contribution to saving patients’ lives