Cell classification with phase-imaging meta-sensors.

Area of Science:

• Optoelectronics
• Machine Learning
• Computational Imaging

Background:

• Photonic technologies offer a path to reduce computational costs in machine learning.
• Convolutional Neural Networks (CNNs) are widely used for image classification.
• Metasurface photodetectors are emerging as novel optical components.

Purpose of the Study:

• To investigate a hybrid optoelectronic CNN architecture for image classification.
• To evaluate the performance of angle-sensitive metasurface photodetectors in a CNN.
• To assess the potential for reducing computational load in image classification.

Main Methods:

• Replacing the first convolutional layer of a CNN with an image sensor array of angle-sensitive metasurface photodetectors.
• Visualizing transparent phase objects by recording anisotropic edge-enhanced images.
• Evaluating classification performance using computational-imaging simulations for cancer cell identification.

Main Results:

• The hybrid optoelectronic network accurately classifies transparent cancer cells (>90% accuracy).
• The system directly visualizes phase objects, mimicking CNN feature maps.
• An order-of-magnitude reduction in calculations was achieved compared to a fully digital CNN.

Conclusions:

• Hybrid optoelectronic networks integrating metasurface photodetectors show promise for efficient image classification.
• This approach offers a significant reduction in computational cost for machine learning tasks.
• The developed system demonstrates high accuracy in identifying transparent biological samples.