Wavefront reconstruction based on deep transfer learning for microscopy.

Area of Science:

• Biomedical imaging
• Machine learning applications
• Optical physics

Background:

• Machine learning (ML) aids real-time, non-invasive deep tissue imaging.
• Biological tissue diversity challenges unified ML model training, reducing accuracy.
• Domain shift between training and testing data limits ML performance in real-world applications.

Purpose of the Study:

• To propose a sensorless wavefront reconstruction method using transfer learning.
• To overcome domain shift issues caused by biological tissue heterogeneity.
• To improve the accuracy and performance of ML models in biomedical imaging.

Main Methods:

• Developed a weights-sharing two-stream convolutional neural network (CNN) framework.
• Utilized labeled simulated data as source-domain and unlabeled specific samples as target-domain data.
• Employed domain adaptation techniques for training on massive labeled simulated data.

Main Results:

• Achieved 18.5% higher accuracy compared to conventional CNN methods.
• Increased peak intensities of the point spread function (PSF) by over 20%.
• Demonstrated comparable training and processing times to existing methods.

Conclusions:

• The proposed transfer learning method effectively addresses domain shift in wavefront reconstruction.
• The approach offers superior performance for complex aberrations in biological tissues.
• This advancement holds significant potential for improving deep tissue imaging in biomedical research.