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

Full abstract

Read the full abstract

The integration of artificial intelligence (AI) with digital holographic microscopy (DHM) is transforming optical methods for particle detection and classification, particularly in biosecurity and biosafety. However, existing AI-DHM methods are typically developed or fine-tuned on hardware-specific experimental datasets, limiting their generalizability across different configurations and applications. A standardized and hardware-agnostic benchmarking framework for systematically evaluating AI performance in DHM is currently lacking. This study introduces a simulation and evaluation framework designed to generate reproducible, parameter-controlled synthetic datasets of raw holographic images, configurable to replicate specific optical setups while remaining independent of any particular hardware implementation. The framework provides a flexible environment for pre-training and testing open-source and proprietary machine learning (ML) and deep learning (DL) models, enabling transfer learning strategies and guiding the design and optimization of optical setups and computational pipelines. Here, the framework is demonstrated for the recognition of micrometric and submicrometric particles relevant to biosecurity scenarios, where particle size and concentration are critical operational parameters that directly influence sampling, filtration strategies and downstream AI analysis. By systematically varying these parameters alongside optical configurations, the framework is used to investigate their individual and combined effects on detection and classification accuracy. Results demonstrate the robustness of DL models, even under challenging conditions with small particles and high concentrations, while ML approaches are more sensitive to fringe overlap. Overall, this work provides a reproducible and extensible environment for the development of AI-driven microscopy systems, offering quantitative insights for optimizing experimental design, algorithm development and sample preparations in both laboratory and field applications, prior to the deployment on specific optical setups.