GPU-based deep convolutional neural network for tomographic phase microscopy with ℓ1 fitting and regularization.

Area of Science:

• Biomedical Imaging
• Computational Microscopy
• Optical Physics

Background:

• Tomographic phase microscopy (TPM) enables 3D refractive index mapping of transparent samples.
• Conventional TPM requires dense angular sampling, limiting temporal resolution and dynamic scene analysis.

Purpose of the Study:

• To develop a graphics processing unit (GPU)-accelerated deep learning method for enhanced TPM.
• To improve TPM performance with significantly reduced angular sampling.

Main Methods:

• Implemented a deep convolutional neural network (CNN) on a GPU for phase tomography.
• Utilized an ℓ1-norm sparsity constraint as a loss function for data fidelity and gradient-domain regularization.
• Employed a multislice beam propagation model.

Main Results:

• Achieved at least 14 dB improvement in signal-to-noise ratio compared to state-of-the-art algorithms.
• Demonstrated successful imaging of HeLa cells with substantial data reduction.
• Validated the effectiveness of the deep learning approach with fewer incident angles.

Conclusions:

• The proposed GPU-accelerated deep learning method significantly enhances TPM performance.
• Reduced angular sampling is feasible for accurate 3D refractive index distribution measurement.
• This advancement broadens the applicability of TPM for dynamic biological imaging.