Deep Learning for Skin Lesion Classification using Pre-Trained Neural Networks

The increasing prevalence of skin cancer underscores the importance of accurate and efficient diagnostic tools. Deep Convolutional Neural Networks (DCNN) have demonstrated remarkable success in computer vision tasks, particularly in skin cancer classification. However, adapting these networks to specific domains, such as dermatoscopic images, poses challenges. This project addresses the need for robust skin cancer classification by fine-tuning DCNNs pre-trained on ImageNet for the specific task of identifying 7 types of skin lesions using the HAM10000 dataset, which contains 10,000 dermatoscopic images. To comprehensively assess the efficiency of different fine-tuning strategies, a comparative analysis is conducted across multiple architectures, including the baseline model, VGG16, Inception V3, InceptionResNetV2, and DenseNet201. The initial examination involves fine-tuning the top layers of each architecture. Subsequently, a more intricate analysis is performed by fine-tuning all layers for Inception V3, InceptionResNetV2, and DenseNet201. This study aimed to improve dermatological image classification by combining cross-learning and ensemble learning. The ensemble included the best versions of Inception V3 and DenseNet 201, achieving an accuracy of 88.52% on the test set and 88.8% on the validation set. The fitting of the whole model improved the results and the speed of convergence.