Abstract
This paper aims to address the pressing issue of melanoma classification by leveraging advanced neural network models, specifically basic Convolutional Neural Networks (CNN), ResNet-18, and EfficientNet-B0. Our objectives encompass presenting and evaluating these models based on established practices in medical image diagnosis. Additionally, we aim to demonstrate their effectiveness in contributing to the critical task of saving lives through early and accurate melanoma diagnosis.Our methodology involves a multi-stage process, which includes image normalization and augmentation, followed by segmentation, feature extraction, and classification. Notably, the neural network models underwent rigorous evaluation, with EfficientNet-B0 exhibiting exceptional performance as the winning model. EfficientNet-B0 achieved a remarkable accuracy of 97%, surpassing ResNet-18 (87%) and basic CNN (80%) in classifying malignant and benign cases. In addition to accuracy, a comprehensive set of evaluation metrics was employed for EfficientNet-B0: sensitivity of 99%, specificity of 93%, F1-score of 97%, precision of 95%, and an error rate of 3%. It also demonstrated a Mathew's correlation coefficient of 94% and a geometric mean of 1.01. Across these metrics, EfficientNet-B0 consistently outperformed ResNet-18 and basic CNN. The findings from this research suggest that neural network models, particularly EfficientNet-B0, hold significant promise for precise and efficient melanoma skin cancer detection.