Deep Learning-based Skin Cancer Diagnosis using Dermoscopy Images

Over the past few years, deep learning-based models have been state-of-the-art for various classification problems in medical imaging. Transfer learning models in deep learning are successfully employed for skin cancer diagnosis. Model scaling is an important parameter in deep learning-based transfer learning models to achieve better classification accuracy. The current work employs the EfficientNet-based model that proposes an efficient model scaling by considering the accuracy of the classification models. Various types of EfficientNet models are finetuned and the penultimate layer features of the models are extracted. The features of the EfficientNet models' dimension are large, and by using the concept of principal component analysis the more significant features are extracted. Feature fusion of the important features is done and classification is performed using the stacked ensemble model. The stacked ensemble model is a two-stage classification model. The stacked ensemble models use kernel-based support vector machine and gradient boosting classifier for the prediction of labels for the dermoscopic images, and logistic regression-based model for classification. The proposed model performance is evaluated for Skin Cancer Detection (SCD) and Skin Cancer Classification (SCC) using the ISIC archive and HAM10000 skin disease dermoscopy datasets. In various test cases, the EfficientNet-based transfer learning approach has shown improved performance of accuracy compared to the other models with an accuracy of 98% in SCD using the ISIC archive dataset and 99% accuracy for SCC using the HAM10000 dataset. The EfficientNet-based finetuned model can serve as an early-stage diagnosis for SCD and SCC in the healthcare environment. The tool enables dermatologists and healthcare professionals to provide proper treatment for patients affected by skin cancer.