Comparative Analysis of ResNet Models for Skin Cancer Diagnosis: Performance Evaluation and Insights

The performance of the ResNet-18, ResNet-34, and ResNet-50 models in identifying skin cancer is evaluated using a dataset of 2,357 photographs obtained from the International Skin Imaging Collaboration (ISIC). The photographs portray a range of skin conditions, including both malignant and benign diseases. The models are evaluated based on their capacity to classify the images into nine unique categories representing various skin issues. The dataset undergoes preprocessing procedures, such as resizing, cropping, and normalizing, prior to its utilization for training and assessment purposes. The Stochastic Gradient Descent (SGD) algorithm is employed for the purpose of model optimization, utilizing a learning rate of 0.001 and a momentum of 0.9. Furthermore, the implementation incorporates a learning rate scheduler known as StepLR, which systematically decreases the learning rate by 0.1 units every 7 epochs over a span of 10 epochs. The CrossEntropyLoss function is employed for the purpose of training the models on the dataset, wherein it quantifies the disparity between the predictor and true class probabilities. The findings indicate that ResNet-50 outperforms the other models, achieving an impressive accuracy of 95.3% and a loss rate of 18.5%. Nevertheless, it is important to acknowledge that these findings are derived from a limited representative sample. In order to enhance the credibility and use of forthcoming research endeavors, it is imperative to augment the magnitude and inclusiveness of the dataset. This study highlights the need of employing sophisticated deep learning models, such as ResNet-50, inside clinical environments to enhance the identification of skin cancer and improve patient outcomes. Furthermore, it emphasizes the significance of continuous study and development in the field of deep learning to enhance medical diagnosis.