Learned sensing: jointly optimized microscope hardware for accurate image classification.

Area of Science:

• Microscopy
• Artificial Intelligence
• Computational Biology

Background:

• Microscopes traditionally rely on human interpretation.
• Deep learning necessitates hardware redesign for automated image analysis.
• Optimizing illumination is crucial for accurate automated classification.

Purpose of the Study:

• To develop a method for co-optimizing microscope illumination and deep neural network-based image classification.
• To enhance the speed and accuracy of automated image analysis for biological samples.
• To design hardware tailored for specific automated interpretation tasks.

Main Methods:

• A deep neural network with an integrated "physical layer" was developed.
• Joint optimization of illumination patterns and classification algorithms was performed.
• The method was tested for identifying malaria-infected cells and diagnosing blood smear types.

Main Results:

• The learned sensing approach for illumination design improved malaria-infected cell identification accuracy by 5-10% compared to standard methods.
• The joint hardware-software design procedure demonstrated generalization across different blood smear types.
• The procedure showed experimental translatability across different setups while maintaining high accuracy.

Conclusions:

• Co-optimizing microscope illumination and deep learning models offers superior automated image analysis.
• This integrated hardware-software approach enhances diagnostic accuracy and generalizability.
• Learned sensing provides a powerful framework for designing task-specific microscopy hardware.