Frequency-Guided Cross-Modal Interaction for Multimodal Yeast Classification Based on Light-Scattering and Microscopy...

Area of Science:

• Microbiology
• Computational Biology
• Biophysics

Background:

• Accurate yeast identification is vital for clinical diagnosis and antifungal treatment, but current microscopy-based methods struggle with generalization.
• Light-scattering (LS) imaging offers volumetric biophysical cues but faces challenges in feature discrimination due to indirect representations.

Purpose of the Study:

• To develop fast and accurate deep learning methods for classifying yeast species using LS and microscopy imaging.
• To enhance yeast classification by leveraging frequency-domain features and multimodal data integration.

Main Methods:

• Proposed FPA-YeastNet, a frequency-enhanced deep learning architecture for single-modality LS image classification.
• Developed FGCA-YeastNet, a frequency-guided cross-attention network integrating LS and microscopy data for complementary representation learning.
• Utilized adaptive fusion and bidirectional attention for synergistic interactions between different imaging modalities.

Main Results:

• FPA-YeastNet improved LS-only model accuracy by an average of 6.26%.
• FGCA-YeastNet achieved mean accuracy gains of 19.97% over unimodal baselines and 7.67% over multimodal baselines.
• Demonstrated FGCA-YeastNet's effectiveness in bridging the performance gap between LS and microscopy modalities.

Conclusions:

• Frequency-guided multimodal collaboration significantly enhances the reliability and interpretability of yeast classification.
• Light scattering and microscopic imaging show diagnostic potential, especially when combined using advanced deep learning techniques.
• The proposed models offer a promising approach for accurate and efficient yeast identification in clinical microbiology settings.