Going Smaller: Attention-based models for automated melanoma diagnosis

Area of Science:

• Artificial Intelligence in Medicine
• Computational Dermatology
• Medical Imaging Analysis

Background:

• Early melanoma diagnosis is crucial for patient outcomes.
• Current advanced diagnostic models are computationally intensive, limiting practical application.
• Large ensemble models with millions of parameters pose challenges for real-world deployment.

Purpose of the Study:

• To develop and evaluate compact attention-based convolutional neural network (CNN) architectures for melanoma classification.
• To assess the feasibility of deploying efficient diagnostic models in resource-limited settings, such as smartphone-based dermoscopy.
• To compare the performance of novel compact models against leading International Skin Imaging Collaboration (ISIC) 2020 challenge winners.

Main Methods:

• Development of novel attention-based CNN architectures using the EfficientNet-B3 backbone.
• Comparative analysis against top International Skin Imaging Collaboration (ISIC) 2020 challenge winners.
• Evaluation using two independent test sets to validate performance and efficiency.

Main Results:

• The proposed compact CNN architectures achieved performance comparable to or better than the top ISIC 2020 challenge winners.
• Models demonstrated high accuracy while utilizing significantly fewer parameters (up to 98% reduction).
• The architectures showed a strong balance between diagnostic efficiency and accuracy.

Conclusions:

• Compact, attention-based CNN models are feasible for accurate melanoma diagnosis.
• These efficient models can revolutionize point-of-care diagnostics, especially in remote or under-resourced areas.
• The developed architectures offer a practical solution for real-time melanoma detection with reduced computational demands.