Multi-Class and Multi-Level Classification for Common Wound Bacteria Based on Reflectance Hyperspectral Imaging.

Hyperspectral imaging is an emerging non-destructive technique for rapid bacterial detection. However, existing methods often inadequately exploit spatial-spectral correlations and lack datasets tailored to clinically relevant scenarios. This work presents a systematic framework integrating a near-infrared hyperspectral imaging system with a deep learning for accurate classification of common wound bacteria. The multi-class and multi-level dataset is first established using a 400-2500 nm halogen-illuminated system under controlled conditions, comprising 26 056 image patches covering seven bacterial species at five concentration levels, totaling 31 classes. To fully harness spatial and spectral information, the Spatial-Spectral Interactive Selection Network (SSIS-Net) is proposed. Evaluated on the self-constructed dataset, SSIS-Net achieves an overall accuracy of 97.40%, an average accuracy of 97.37%, and a Kappa coefficient of 97.31%, consistently outperforming state-of-the-art methods. These results demonstrate that the proposed system offers a reliable, rapid, and non-destructive solution for bacterial detection, with strong potential for clinical integration.