Transfer Learning-Based Skin Cancer Classification Using a Modified VGG19 Network

Abstract

Skin cancer continues to be one of the most common and potentially fatal types of cancer globally, with the need for
early and precise diagnosis to enhance patient outcomes. This article introduces a novel method of automated skin cancer
classification based on a Modified VGG19 deep learning model. Embracing the philosophy of transfer learning, the model utilizes
pre-trained weights of ImageNet with selective fine-tuning of the last layers to better fit the unique characteristics of skin cancer
images. Also, a dual transfer learning approach was utilized with retraining the early layers of a pre-trained model AlexNet to
improve detection of lesion boundaries. The model was trained and tested on the publicly available HAM10000 dermoscopic image
database with 5-fold cross-validation. The model with the suggested approach attained superior performance with an average
validation accuracy of 99.07%. The proposed technique proved to be more accurate and robust when compared to other current
existing classification methods. The results imply that Modified VGG19 model, augmented with an adapted transfer learning
framework, is exceedingly suitable for the diagnosis of skin cancer and is strongly promising for real-world clinical use.