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

Area of Science:

• Biophotonics
• Medical Imaging
• Machine Learning

Background:

• Hyperspectral imaging (HSI) offers non-destructive bacterial detection but struggles with spatial-spectral correlations and clinical datasets.
• Existing methods lack comprehensive datasets for clinically relevant bacterial identification scenarios.

Purpose of the Study:

• To develop a robust framework integrating near-infrared HSI and deep learning for accurate wound bacteria classification.
• To create a tailored, multi-class, multi-level dataset for training and evaluating bacterial detection models.

Main Methods:

• A near-infrared HSI system (400-2500 nm) was used to capture images under controlled conditions.
• A dataset of 26,056 image patches covering seven bacterial species at five concentrations (31 classes) was established.
• The Spatial-Spectral Interactive Selection Network (SSIS-Net) was proposed to leverage spatial and spectral information.

Main Results:

• SSIS-Net achieved high performance on the self-constructed dataset, with 97.40% overall accuracy, 97.37% average accuracy, and 97.31% Kappa coefficient.
• The proposed method demonstrated superior performance compared to existing state-of-the-art techniques.
• The developed system provides reliable, rapid, and non-destructive bacterial detection capabilities.

Conclusions:

• The integrated HSI and deep learning system offers a promising solution for bacterial detection.
• The SSIS-Net model effectively utilizes spatial-spectral information for accurate classification.
• The system shows significant potential for integration into clinical settings for wound management.