A wavelet-guided transformer approach for autofocus in brightfield biological microscopy

Area of Science:

• Microscopy
• Artificial Intelligence
• Image Processing

Background:

• Autofocus is critical for biological microscopy image clarity and efficiency.
• Conventional autofocus methods in brightfield microscopy have limitations in real-time performance and noise sensitivity.
• These limitations hinder applications requiring rapid, precise focusing.

Purpose of the Study:

• To develop a novel autofocus method for brightfield biological microscopy.
• To improve real-time performance and accuracy compared to existing techniques.
• To enable high-throughput microscopy applications.

Main Methods:

• Proposed a Wavelet-Guided Transformer Network (WGT-Net) for autofocus prediction from single blurred images.
• Integrated wavelet transform for multi-scale feature extraction and downsampling.
• Utilized a Transformer module for global-local dependency analysis and a Gaussian soft labeling strategy for uncertainty handling.

Main Results:

• WGT-Net achieved a Mean Absolute Error (MAE) of 0.0869 and Root Mean Square Error (RMSE) of 0.101.
• Demonstrated significant reductions in MAE (28.69%) and RMSE (32.39%) compared to state-of-the-art methods.
• Achieved autofocus predictions within milliseconds, showcasing exceptional real-time capability.

Conclusions:

• WGT-Net offers substantial improvements in autofocus accuracy and real-time performance for biological microscopy.
• The method is highly suitable for real-time, high-throughput brightfield biological microscopy.
• This advancement addresses critical limitations of conventional autofocus systems.