Integration of Feature Fusion Strategy on EfficientNet for Skin Cancer Detection and Classification

Abstract

Skin cancer is a disorder that arises from changes in healthy skin cells that give them the ability to become malignant. Due to a rise in predominance over the past ten years, it is currently placed among the top ten malignancies in terms of frequency. Patients who are unaware of skin cancer may not be encouraged to seek medical attention for minor skin discoloration because many people lack the knowledge required to notice it. One can lessen and manage the detrimental effects of skin cancer with an accurate diagnosis and prompt, efficient therapy. Investigating skin cancer lesions can be difficult due to their comparatively similar forms, complex expression of the disease, and susceptibility to subjective diagnosis. The obtained features from the multiple EfficientNet model tiers are combined using a feature fusion approach to solve this challenge. Therefore, in this study, a system was developed that could detect and classify skin cancer lesions into benign and malignant automatically by using a feature fusion strategy and EfficientNet algorithm with a transfer learning method. The image dataset was collected from a public dataset that is available on Kaggle and the total dataset used is 27560 from both classes benign and malignant. Pre-processing the skin lesions, extracting features using a pre-trained EfficientNet, feature concatenation, and classifying using deep learning EfficientNet algorithm are the primary components of this research. The study method was tested and yielded average results of 93.4% accuracy, 92.3% precision, 94.8% recall, 92.1% specificity, and 93.5% f1-score, respectively as well as confusion matrix achieved 1269(92.00%) true positives, 109(8.00%) false positives, 72(5.00%) false negatives, and 1306(95.00%) true negatives