Convolutional Neural Network Transformer (CNNT) for Fluorescence Microscopy image Denoising with Improved...

Area of Science:

• Microscopy
• Image Processing
• Deep Learning

Background:

• Deep neural networks improve fluorescence microscopy image quality.
• Convolutional neural networks (CNNs) require extensive, experiment-specific training.
• Existing methods lack generalizability and broad applicability.

Purpose of the Study:

• Introduce a novel Convolutional Neural Network Transformer (CNNT) model.
• Develop a more efficient and adaptable deep learning approach for image denoising.
• Improve image quality in fluorescence microscopy while reducing training time.

Main Methods:

• Trained a single CNNT backbone model using pairwise high-low signal-to-noise ratio (SNR) images.
• Utilized a fine-tuning strategy with 5-10 sample pairs for new applications.
• Evaluated CNNT performance against CNN-based methods like RCAN and Noise2Fast.

Main Results:

• CNNT significantly reduced training time compared to separate CNN models.
• Achieved superior image denoising performance across various microscopy techniques.
• Demonstrated fast adaptation to new imaging experiments through fine-tuning.
• Reduced confocal microscopy scan time from one hour to eight minutes with enhanced quality.

Conclusions:

• The CNNT backbone and fine-tuning scheme offers a powerful and efficient solution for fluorescence microscopy image enhancement.
• This approach overcomes the limitations of traditional CNNs, improving applicability and generalization.
• CNNT enables faster, higher-quality imaging across diverse fluorescence microscopy applications.