TLDR: RT-HAD is a new deep learning system that uses digital holographic microscopy to quickly and accurately detect blood cell aggregates, like platelet clumps, in real-time. It processes large amounts of data on-the-fly, significantly reduces data storage needs, and provides rapid diagnostic insights for conditions such as sepsis, offering a major improvement for point-of-care haematology.
The field of haematology diagnostics is on the cusp of a significant advancement with the introduction of RT-HAD, a novel deep learning-based system designed to identify crucial blood cell aggregates in real-time. This innovation promises to overcome long-standing challenges in diagnosing conditions like sepsis and thrombo-inflammatory disorders, where these aggregates serve as vital early biomarkers.
Traditional methods for analyzing blood, such as conventional flow cytometers, are excellent for counting cells but often miss or misidentify complex blood cell aggregates. This leads to “flagged” results that require time-consuming manual review by experts. While advanced techniques like imaging flow cytometry (IFC) can capture detailed aggregate morphologies, their clinical use is limited by extensive sample preparation, slow data acquisition, and massive data storage requirements. A single patient’s data can easily exceed 30 GB, posing a “big data” problem for routine clinical application.
RT-HAD, which stands for Real-Time Holographic Aggregate Detector, addresses these limitations by integrating off-axis digital holographic microscopy (DHM) with an end-to-end deep learning framework. DHM is a label-free imaging technique that captures rich morphological data of blood cells without the need for dyes or extensive sample preparation, making it ideal for high-speed flow systems. However, the raw data from DHM is computationally intensive to process.
How RT-HAD Works
The RT-HAD system is composed of three specialized deep learning modules working in harmony. First, it uses OAH-Net, a physics-consistent holographic reconstruction model, to quickly convert raw holographic images into high-resolution phase and amplitude images. This step is crucial because directly analyzing raw holograms is much slower and less accurate for detecting small features like platelets. OAH-Net can process a frame in just 4.7 milliseconds.
Second, a deep learning-based object detection model, specifically a variant of YOLOv8x-p2, identifies and classifies individual blood cells, including erythrocytes, leukocytes, and platelets, from the reconstructed phase images. This module is highly accurate, especially for small platelets, achieving over 96% precision. It processes each frame in about 6.6 milliseconds.
Finally, a graph-based aggregate analyzer takes the detected individual cells and represents them as nodes in a graph. By analyzing their spatial relationships, it identifies and categorizes blood cell aggregates, such as platelet-platelet (PP), leukocyte-leukocyte (LL), and leukocyte-platelet (LP) interactions. This module is very efficient, adding only about 0.5 milliseconds per frame.
Also Read:
- RedDino: A New AI Foundation Model for Advanced Red Blood Cell Analysis
- Enhancing Sepsis Treatment with Multimodal AI: A New Approach Leveraging Clinical Notes
Impact and Benefits
The combined power of these modules allows RT-HAD to process over 30 GB of image data on-the-fly, delivering results in less than 1.5 minutes. This turnaround time is significantly faster than previous DHM processing methods, which could take hours, and meets the requirements for point-of-care testing in acute care settings. Furthermore, RT-HAD boasts an error rate of only 8.9% in platelet aggregate detection compared to human experts, which falls within acceptable laboratory error rates for haematology biomarkers.
A major advantage of RT-HAD is its ability to drastically reduce data storage needs. Instead of storing massive raw holographic frames, the system only saves relevant regions of interest (ROIs) containing detected cells and aggregates. This reduces storage size by over 99%, transforming 30 GB of raw data into approximately 15 MB of analysis results per patient. This not only offers substantial economic and engineering benefits but also reduces the carbon footprint associated with data storage.
The clinical utility of RT-HAD has been demonstrated through patient case studies, where it successfully quantified platelet microaggregates as potential early predictive biomarkers. For instance, in pneumonia patients, changes in platelet aggregate levels and sizes correlated well with disease progression and recovery, providing earlier and more detailed insights than traditional platelet counts or SOFA scores. This suggests that RT-HAD can enhance routine haematology panels by revealing “hidden” biomarkers without interrupting the clinical workflow.
In essence, RT-HAD represents a significant leap forward in real-time haematology imaging. It offers a scalable, cost-effective, and sustainable diagnostic platform that overcomes key limitations of existing technologies, paving the way for more precise and timely diagnostics in acute care. For more detailed information, you can refer to the full research paper available here.
