A simulated dataset and evaluation framework for assessing AI detection limits in digital holographic microscopy.

Area of Science:

• Optical microscopy
• Artificial intelligence
• Biosecurity

Background:

• Artificial intelligence (AI) integrated with digital holographic microscopy (DHM) shows promise for particle detection and classification in biosecurity.
• Current AI-DHM methods lack generalizability due to hardware-specific datasets, hindering widespread application.
• A standardized, hardware-agnostic benchmarking framework for AI in DHM is needed.

Purpose of the Study:

• To introduce a simulation and evaluation framework for generating reproducible, synthetic DHM datasets.
• To create a hardware-agnostic environment for testing and optimizing AI models in DHM.
• To investigate the impact of particle characteristics and optical configurations on AI performance.

Main Methods:

• Developed a simulation framework to generate synthetic raw holographic images with controlled parameters.
• Configured the framework to mimic diverse optical setups independently of specific hardware.
• Utilized the framework to pre-train and test machine learning (ML) and deep learning (DL) models for particle recognition.

Main Results:

• Demonstrated the framework's capability to generate parameter-controlled synthetic datasets for AI-DHM development.
• Showcased the robustness of deep learning models in detecting small particles at high concentrations.
• Identified that ML approaches are more susceptible to fringe overlap issues compared to DL models.

Conclusions:

• The developed framework provides a reproducible and extensible environment for AI-driven microscopy.
• Offers quantitative insights for optimizing experimental design, algorithm development, and sample preparation for DHM systems.
• Facilitates the development and deployment of AI-DHM for biosecurity and biosafety applications.