Scale-Adaptive Deep Model for Bacterial Raman Spectra Identification.

Area of Science:

• Analytical Chemistry
• Biotechnology
• Computational Biology

Background:

• Raman spectroscopy combined with deep learning offers rapid bacterial diagnosis.
• Current deep learning models have accuracy limitations due to fixed feature scales.

Purpose of the Study:

• To develop a deep neural network capable of learning multi-scale features from Raman spectra for improved bacterial identification.
• To enhance model interpretability by visualizing key spectral features.

Main Methods:

• Proposed a novel deep neural network architecture incorporating multi-receptive fields in convolutional layers.
• Utilized expert knowledge on multi-scale spectral peaks for feature discrimination.
• Visualized activated wavenumbers to enhance model interpretability.

Main Results:

• Achieved superior accuracy and efficiency in bacterial identification across isolate, empiric-treatment, and antibiotic-resistance levels.
• Demonstrated significantly reduced sample requirements for clinical bacterial identification tasks.
• Outperformed existing state-of-the-art methods in bacterial spectrum analysis.

Conclusions:

• The proposed multi-scale deep learning model significantly advances automated bacterial identification.
• This method holds substantial potential for clinical diagnostics, antibiotic susceptibility testing, and personalized prescription guidance.