Enhancing Skin Cancer Diagnosis Through Fine-Tuning of Pretrained Models: A Two-Phase Transfer Learning Approach.

Abstract

Skin cancer is among the most prevalent types of cancer worldwide, and early detection is crucial for improving treatment outcomes and patient survival rates. Traditional diagnostic methods, often reliant on visual examination and manual evaluation, can be subjective and time-consuming, leading to variability in accuracy. Recent developments in machine learning, particularly using pretrained models and fine-tuning techniques, offer promising advancements in automating and improving skin cancer classification. This paper explores the application of a two-phase model using the HAM10000 dataset, which comprises a wide range of skin lesion images. The first phase employs transfer learning with frozen layers, followed by fine-tuning all layers in the second phase to adapt the models more specifically to the dataset. I evaluate nine pretrained models, including VGG16, VGG19, InceptionV3, Xception (extreme inception), and DenseNet121, assessing their performance based on accuracy, precision, recall, and F1 score metrics. The VGG16 model, after fine-tuning, achieved the highest test set accuracy of 99.3%, highlighting its potential for highly accurate skin cancer classification. This study provides important insights for clinicians and researchers, demonstrating the efficacy of advanced machine learning models in enhancing diagnostic accuracy and supporting clinical decision-making in dermatology.