Real-Time Control and Automation Framework for Acousto-Holographic Microscopy

TL;DR

This work addresses the bottleneck of manual microscopy in cell biology by delivering a fully automated DHM platform for autonomous mechanical characterization. It couples automated serpentine scanning, real-time YOLO-based detection, a robust holographic autofocus pipeline, and GPU-accelerated parallel reconstruction to process 250-frame hologram sequences at 50 fps. Key contributions include a multimodal autofocus metric robust to noisy holograms, a data-driven throughput model that identifies the 2.23 s autofocus as the primary bottleneck, and validation of autonomous screening capabilities. The authors also discuss future enhancements via hybrid brightfield imaging and advanced segmentation to enable scalable DHM-based screening in biology.

Abstract

Manual operation of microscopes for repetitive tasks in cell biology is a significant bottleneck, consuming invaluable expert time, and introducing human error. Automation is essential, and while Digital Holographic Microscopy (DHM) offers powerful, label-free quantitative phase imaging (QPI), its inherently noisy and low-contrast holograms make robust autofocus and object detection challenging. We present the design, integration, and validation of a fully automated closed-loop DHM system engineered for high-throughput mechanical characterization of biological cells. The system integrates automated serpentine scanning, real-time YOLO-based object detection, and a high-performance, multi-threaded software architecture using pinned memory and SPSC queues. This design enables the GPU-accelerated reconstruction pipeline to run fully in parallel with the 50 fps data acquisition, adding no sequential overhead. A key contribution is the validation of a robust, multi-stage holographic autofocus strategy; we demonstrate that a selected metric (based on a low-pass filter and standard deviation) provides reliable focusing for noisy holograms where conventional methods (e.g., Tenengrad, Laplacian) fail entirely. Performance analysis of the complete system identifies the 2.23-second autofocus operation-not reconstruction-as the primary throughput bottleneck, resulting in a 9.62-second analysis time per object. This work delivers a complete functional platform for autonomous DHM screening and provides a clear, data-driven path for future optimization, proposing a hybrid brightfield imaging modality to address current bottlenecks.

Figures (7)

• Figure 1: Side-view of the Mach-Zehnder Interferometer Setup
• Figure 2: Process Chart for the Real-Time Image Processing and Control Subsystem.
• Figure 3: Data Flow Diagram for the Real-Time Reconstruction and Quantification Subsystem
• Figure 4: Training and evaluation metrics of the YOLO model across 100 epochs.
• Figure 5: Sample inference result showing ground truth and predicted bounding boxes. True positives, false positives and false negatives are separately remarked in the results.