Enhanced skin cancer classification using modified efficientNetV2L with adaptive early stopping mechanism.

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate skin cancer classification is vital for medical diagnostics.
• Deep learning models, particularly Convolutional Neural Networks (CNNs), are used for skin lesion analysis.
• Existing CNN models struggle with generalization, overfitting, and imbalanced datasets, limiting clinical application.

Purpose of the Study:

• To develop a robust and efficient deep learning model for accurate skin lesion classification.
• To address the limitations of traditional CNNs in handling diverse and imbalanced skin lesion datasets.
• To improve the generalization capability of models for unseen data in skin cancer diagnosis.

Main Methods:

• Proposed a novel model utilizing the EfficientNetV2L architecture for skin lesion classification.
• Implemented adaptive early stopping and learning rate callbacks to enhance model generalization and prevent overfitting.
• Trained and evaluated the model on the publicly available International Skin Imaging Collaboration (ISIC) dataset.

Main Results:

• The EfficientNetV2L-based model achieved a high classification accuracy of 99.22% on the ISIC dataset.
• Demonstrated robustness in classifying various types of skin lesions.
• The adaptive callbacks effectively improved generalization and mitigated overfitting issues.

Conclusions:

• The developed EfficientNetV2L model offers a powerful and efficient solution for automated skin cancer diagnosis.
• The proposed approach enhances model reliability and clinical relevance for distinguishing between skin lesions.
• This work contributes to advancing the field of AI-driven medical diagnostics in dermatology.