Instance segmentation from particle holograms

Abstract Holographic microscopy has emerged as a low-cost and highly compact technique for 3D imaging of microscopic particles in suspension. However, its broad application is largely limited by the inclusion of multiple steps in extracting the particles from the hologram image, which can be computationally expensive and often involves human intervention. We introduce HoloDINO, a transformative model that leverages instance segmentation for streamlined, end-to-end particle detection and contour extraction. By pre-training a data-intensive transformer model on synthetic particle contours and fine-tuning it on experimental data, our approach demonstrates robust performance across synthetic holograms of varying particle concentration, morphology, and optical properties, as well as experimental holograms of dental aerosols and water spray droplets. HoloDINO surpasses conventional methods, which typically involve multiple steps—such as reconstruction, autofocusing, and segmentation—by consolidating these into a single, efficient process that delivers precise morphological data for each particle in one forward pass. This advancement not only facilitates real-time applications but also significantly enhances the generalization capabilities across diverse settings, paving the way for broader adoption of holography in particle analysis.