A brain-inspired computational framework for image-based risk assessment

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Computational Biology

Background:

• Skin cancer risk prediction is crucial for early diagnosis and personalized healthcare.
• Existing methods face challenges in balancing feature representation richness and computational efficiency.
• There is a need for advanced computational frameworks to improve skin cancer risk assessment.

Purpose of the Study:

• To propose Bicom, a brain-inspired framework for enhanced skin cancer risk prediction.
• To integrate efficient attention mechanisms, multi-scale feature fusion, and confidence-aware refinement.
• To improve the accuracy, robustness, and scalability of assisted skin cancer risk prediction.

Main Methods:

• Developed F-ResNeSt, a multi-scale feature extractor using Feature Pyramid Network (FPN) and linear-complexity attention (Linformer).
• Proposed L-CoAtNet, an optimized classification network employing Linformer-based attention for scalable global contextual modeling.
• Introduced a Spiking Neural Network (SNN) module for confidence-aware refinement of predictions on ambiguous samples.

Main Results:

• The Bicom framework demonstrated consistently competitive performance across multiple evaluation metrics.
• Experiments were conducted on both public and subject-specific datasets.
• The proposed methods achieved significant improvements in skin cancer risk prediction accuracy and reliability.

Conclusions:

• The Bicom framework offers an effective, robust, and scalable solution for assisted skin cancer risk prediction.
• The integration of advanced deep learning techniques and brain-inspired modules enhances prediction reliability.
• This approach holds promise for advancing early diagnosis and personalized healthcare in dermatology.