MDF2Former: Multi-Scale Dual-Domain Feature Fusion Transformer for Hyperspectral Image Classification of Bacteria in...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Microbiology

Background:

• Bacterial wound infections are a significant clinical challenge, potentially leading to sepsis and organ failure.
• Current pathogen identification methods are often complex and time-consuming, hindering timely treatment.
• Rapid and accurate bacterial identification is crucial for effective wound infection management.

Purpose of the Study:

• To develop a hyperspectral imaging (HSI) system for bacterial analysis.
• To create a deep learning model, MDF2Former, for classifying wound bacteria using HSI data.
• To evaluate the performance of the developed system and model in identifying bacterial species.

Main Methods:

• Development of a hyperspectral imaging (HSI) acquisition system tailored for bacterial analysis.
• Implementation of a Multi-Scale Dual-Domain Feature Fusion Transformer (MDF2Former) model.
• MDF2Former incorporates multi-scale feature fusion, spatial-spectral attention, and hierarchical encoding for efficient feature learning.

Main Results:

• The MDF2Former model achieved high performance metrics on a self-constructed HSI dataset of wound bacteria.
• Key performance indicators included Accuracy (91.94%), Precision (92.26%), Recall (91.94%), F1-score (92.01%), and Kappa coefficient (90.73%).
• The model significantly outperformed existing comparative methods in bacterial classification.

Conclusions:

• Combining HSI with deep learning, specifically MDF2Former, is effective for bacterial identification in wound infections.
• This approach shows potential for assisting in bacterial species identification and informing personalized treatment decisions.
• The developed HSI-based system offers a promising advancement for clinical wound care and infection management.