Using deep learning architectures for skin cancer classification

Skin cancer, one of the most common types of cancer, requires early and accurate diagnosis for successful treatment. For this reason, computer vision systems are used to assist specialist doctors. Especially with the development of deep learning architectures, this process has accelerated even more. In recent years, Convolutional Neural Networks (CNNs) have become a powerful deep learning method that has shown impressive results in image-based classification applications. For this reason, in this thesis, the original versions of the popular CNN models ResNet, InceptionV3, AlexNet and VGG16 and the adapted models obtained by modifying these models are used for skin cancer classification. In addition to these models, a new novel CNN model was developed. All these deep learning models were tested on the ISIC dataset, which is widely recognized for its comprehensive images of skin cancer. Before using this dataset, preprocessing was performed on the dataset as part of the thesis. With this preprocessing, the imbalance in the amount of data between classes was eliminated with data augmentation techniques. Deep learning models were trained with the reorganized dataset and the performance of the models was measured on the test data. The training of the models was performed by rearranging the image sizes in the dataset according to the size in the input layer of the deep learning models. The proposed CNN model achieved an accuracy of 83.22% for the dataset with an image size of 224x224x3 and a higher accuracy of 86% for images with a size of 128x128x3. These results demonstrate the versatility of the proposed CNN model and its potential for practical clinical applications by demonstrating its strong performance at various image resolutions. The results obtained with the proposed CNN model are extensively compared with other developed and pre-trained CNN models. Since the pre-trained models have been previously trained on high dimensional datasets, they have shown relatively high performance compared to the proposed method in this study. The VGG16 model, which is one of the pre-trained models and shows the lowest accuracy, performed relatively poorly with an accuracy of 86%. On the other hand, InceptionV3 outperformed the other models with a high accuracy of 90%, while AlexNet showed a remarkable accuracy of 89%. Finally, ResNet50 performed well with an accuracy of 88%. The results obtained show that the proposed model achieved an accuracy rate close to these models despite being trained with fewer datasets than the pre-trained models. This thesis study emphasized the importance of utilizing advanced deep learning models and comprehensive datasets to improve diagnostic accuracy and thus achieve better clinical outcomes in skin cancer treatment.