Utilizing Transfer Learning for Enhanced Classification of Skin Lesions Through Deep Learning Approaches

The early detection of skin diseases holds great importance in saving lives and reducing the burden of costly treatments. However, dermatologists face a challenge in efficiently managing diagnoses due to the growing prevalence of these diseases. In response, significant advancements have been made in medical image processing techniques, specifically through the implementation of computer-aided diagnosis systems utilizing machine learning and deep learning. Despite this progress, the availability of annotated training data remains limited, posing a major obstacle when training complex and effective classification models. This often leads to overfitting, even when augmented datasets are used. This study proposes a strategy to address these challenges in the classification of skin diseases by employing transfer learning and a customized model structure. The objective is to surpass existing methodologies on the widely recognized HAM10000 dataset. The effectiveness of transfer learning is examined using well-established pre-trained convolutional neural network (CNN) structures, such as ResNet50, MobileNetV2, Xception, InceptionV3, VGG16, and DenseNet121, as base models. Experimental results have revealed that DenseNet121 and MobileNetV2 demonstrate exceptional performance compared to the other architectures investigated, including the ones mentioned in existing literature.