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

The global burden of acute and chronic wounds continues to demand improved classification methods, a critical step in guiding appropriate treatment plans. In this work, we present a lightweight dual-backbone deep learning architecture that combines EfficientNet-B3 and MobileNetV3-Small in a parallel feature extraction framework for wound image classification. The proposed network employs depthwise separable reduction layers, dynamic multi-scale mixers, parallel squeeze-and-excitation attention modules, efficient channel attention blocks, and a learned gated fusion mechanism to merge complementary features from both backbones. Our model was trained and evaluated on two publicly available datasets (AZH and Medetec) across whole image and Region of Interest (ROI) classification protocols using a 70/15/15 train-validation-test split, with all metrics reported as mean ± standard deviation across 10 independent trials. On the four-class whole image AZH benchmark, the proposed model achieves 84.46 ± 2.30% accuracy, improving on the previous best of 83.13%. On the six-class ROI task it reaches 85.83 ± 0.59% accuracy, competitive with prior work while using image data alone without requiring wound location metadata. Grad-CAM visualizations confirm that the network focuses on clinically relevant wound regions, supporting its potential as a decision-support tool for wound care practitioners.