Skin cancer classification based on an optimized convolutional neural network and multicriteria decision-making

Area of Science:

• Dermatology and Artificial Intelligence
• Computational Pathology
• Medical Image Analysis

Background:

• Early detection of skin cancer is crucial for effective treatment.
• Numerous artificial intelligence models exist for skin cancer detection, but optimal model selection is often overlooked.
• Benchmarking various models is essential for advancing AI in skin cancer diagnostics.

Purpose of the Study:

• To develop and compare multiple artificial intelligence models for skin cancer classification.
• To identify the optimal model for skin cancer classification through a systematic benchmarking process.
• To evaluate the impact of feature reduction techniques on model performance.

Main Methods:

• Utilized four convolutional neural network (CNN) architectures (AlexNet, Inception V3, MobileNet V2, ResNet 50) for feature extraction.
• Implemented feature reduction using the grey wolf optimizer (GWO) algorithm and compared with original features.
• Classified skin cancer images into four classes using six machine learning (ML) classifiers, resulting in 51 distinct models.
• Employed the RAPS (ranking alternatives by perimeter similarity) multicriteria decision-making approach for model selection.
• Trained and tested models on the International Skin Imaging Collaboration (ISIC) 2017 dataset.

Main Results:

• The AlexNet CNN combined with the classical GWO algorithm for feature reduction emerged as the optimal model.
• This optimal model achieved a high classification accuracy of 94.5% on the ISIC 2017 dataset.
• Feature reduction using GWO demonstrated benefits in reducing training time and enhancing classification accuracy.
• The RAPS method proved effective in robustly selecting the best-performing model among the 51 developed.

Conclusions:

• This study provides a comprehensive benchmark for skin cancer classification models.
• The optimal model, AlexNet with GWO, offers a promising approach for accurate and efficient skin cancer detection.
• Feature reduction techniques are vital for improving AI model performance in medical image analysis.
• The RAPS method is a valuable tool for selecting optimal AI models in complex classification tasks.