Lightweight dual-backbone network with attentional fusion for wound image classification

Area of Science:

• Medical imaging
• Artificial intelligence in healthcare
• Wound care technology

Background:

• The global burden of acute and chronic wounds necessitates advanced classification methods for effective treatment.
• Current wound classification systems require improvement to guide clinical decision-making accurately.

Purpose of the Study:

• To develop and evaluate a lightweight dual-backbone deep learning architecture for wound image classification.
• To enhance the accuracy and efficiency of wound classification using artificial intelligence.

Main Methods:

• A novel deep learning architecture combining EfficientNet-B3 and MobileNetV3-Small was designed for parallel feature extraction.
• The model incorporates depthwise separable reduction layers, dynamic multi-scale mixers, and attention modules (squeeze-and-excitation, efficient channel attention).
• Training and evaluation were performed on AZH and Medetec datasets using whole image and Region of Interest (ROI) classification protocols.

Main Results:

• The proposed model achieved 84.46 ± 2.30% accuracy on the four-class whole image AZH benchmark, surpassing previous results.
• For the six-class ROI task, the model reached 85.83 ± 0.59% accuracy, utilizing only image data.
• Grad-CAM visualizations demonstrated the network's focus on clinically relevant wound regions.

Conclusions:

• The lightweight dual-backbone deep learning model shows significant potential for accurate wound image classification.
• This AI-driven approach can serve as a valuable decision-support tool for wound care practitioners.
• The model's ability to focus on relevant wound areas without metadata highlights its practical applicability.