Explainable depth-wise and channel-wise fusion models for multi-class skin lesion classification.

Area of Science:

• Artificial Intelligence
• Dermatology
• Medical Imaging

Background:

• Clinical adoption of deep learning in dermatology necessitates models that are both accurate and trustworthy.
• Current AI models often lack transparency, hindering clinical integration.

Purpose of the Study:

• To systematically investigate deep feature fusion for combining complementary representations from diverse neural network architectures.
• To develop high-performing, transparent, and clinically reliable AI diagnostic tools for dermatology.

Main Methods:

• Designed and evaluated six distinct fusion models integrating Convolutional Neural Network (CNN) backbones with Vision Transformers (ViTs).
• Employed depth-wise and channel-wise strategies for feature integration.
• Utilized Grad-CAM and SHAP for explainable AI (XAI) analysis.

Main Results:

• The optimized fusion architecture achieved a 90% weighted average Precision, Recall, and F1 score on the HAM10000 dataset.
• Explainable AI analysis demonstrated that the fusion strategy enhances clinical interpretability.
• Models learned to focus on clinically relevant dermatological features like border irregularity and color variegation.

Conclusions:

• Deep feature fusion offers a robust framework for developing next-generation AI diagnostic tools in dermatology.
• The fusion strategy significantly impacts model interpretability, aligning AI reasoning with expert dermatological knowledge.
• This approach facilitates the creation of transparent and trustworthy AI for clinical use.