A Physics-Inspired Deep Learning Framework for an Efficient Fourier Ptychographic Microscopy Reconstruction under Low...

Area of Science:

• Computational imaging
• Microscopy
• Biomedical imaging

Background:

• High-resolution 2D imaging of biological samples is crucial for medical applications.
• Fourier Ptychographic Microscopy (FPM) offers super-resolution but typically requires numerous raw images (N=225) for reconstruction.
• Reducing the number of raw images is essential to increase microscope throughput.

Purpose of the Study:

• To develop an efficient FPM image reconstruction method using significantly fewer raw images (N=37).
• To enhance microscope throughput without sacrificing image quality or resolution.
• To introduce a novel algorithmic approach combining deep learning and physics-informed optimization.

Main Methods:

• Developed a physics-informed deep neural network for FPM image reconstruction.
• Integrated statistical reconstruction learning for network initialization.
• Explicitly incorporated the forward microscope image formation model into the neural network architecture.

Main Results:

• Successfully reconstructed high-quality 2D images using only 37 raw images.
• Demonstrated no appreciable degradation in image resolution compared to traditional methods.
• Validated the effectiveness of the physics-informed deep neural network approach through simulations.

Conclusions:

• The proposed deep neural network method enables efficient FPM image reconstruction with reduced data acquisition.
• This approach significantly increases microscope throughput while maintaining high image fidelity and resolution.
• The mandatory learning step is critical for achieving optimal reconstruction results.