HoloHema: Digital Holographic Hematology Analyzer

This industrial Ph.D. project, carried out in collaboration between Radiometer Medical ApS and SDU Centre for Photonics Engineering at the University of Southern Denmark, explored the use of digital holographic microscopy (DHM) for the purposes of differential white blood cell counts (dWBCs) in point-of-care (PoC) devices for acute care settings. Two DHM prototypes were developed; an initial lens-based system serving as the foundation for algorithm development, and experimental validation of the approach, achieving 89.6% classification accuracy on a 3-part differential, and a subsequent lensless system for simplified design and increased field-of-view (FoV). Both prototypes employed convolutional neural networks (CNNs) for cell classification. With further optimizations, the lensless system achieved classification accuracies of 92.65% and 89.44% on the 3-part and 5-part differential, respectively. With the lensless system, the derivation of the monocyte distribution width (MDW), a biomarker for sepsis, was also demonstrated. Additionally, pixel super-resolution and multi-wavelength DHM approaches were investigated to enhance the obtained cell information. Finally, a proof-of-principle physics-informed neural network (PINN) for holographic reconstruction was implemented, demonstrating the potential for machine learning (ML) reconstruction techniques. In summary, this work represents an initial exploration of DHM for dWBC in PoC devices, laying the groundwork for future research.

💡 Research Summary

This dissertation reports a comprehensive industrial Ph.D. project carried out jointly by Radiometer Medical ApS and the Centre for Photonics Engineering at the University of Southern Denmark, aiming to create a point‑of‑care (PoC) device capable of performing differential white blood cell counts (dWBC) using digital holographic microscopy (DHM). The work is organized around the design, implementation, and validation of two successive DHM prototypes, followed by a series of auxiliary investigations (pixel super‑resolution, multi‑wavelength imaging, and physics‑informed neural network reconstruction) that together lay the groundwork for a practical PoC hematology analyzer.

Lens‑based DHM prototype
The first system employs a conventional imaging lens to record on‑axis holograms of thin blood smears. The recorded intensity patterns are numerically propagated using the Angular Spectrum Method, and phase retrieval is performed with Gerchberg‑Saxton iterations to obtain high‑quality complex‑field images. These images serve as training data for a convolutional neural network (CNN) that classifies white blood cells into three morphological groups (neutrophils, lymphocytes, monocytes). The lens‑based platform achieves 89.6 % accuracy on a three‑part differential, comparable to bright‑field microscopy‑based automated counters, while providing quantitative phase information that is unavailable in conventional imaging.

Lens‑less DHM prototype
To reduce optical complexity, cost, and to increase the field‑of‑view (FoV), a second “lens‑less” architecture is introduced. Here a coherent laser source illuminates the sample directly onto a CMOS sensor, and the resulting diffraction pattern is recorded without any intervening optics. After background subtraction and digital phase correction, the raw hologram is fed into an optimized CNN. With additional data‑augmentation, network pruning, and hyper‑parameter tuning, the lens‑less system reaches 92.65 % accuracy for three‑part and 89.44 % for five‑part differentials (adding eosinophils and basophils). Importantly, the same data stream is used to compute the Monocyte Distribution Width (MDW), a biomarker linked to sepsis, demonstrating that the device can deliver both cell counts and a clinically relevant functional metric.

Auxiliary technical developments

1. Pixel Super‑Resolution (SR) – Multiple sub‑pixel shifted holograms are combined using an iterative SR algorithm, effectively increasing the sampling frequency beyond the native sensor pixel pitch. This yields sharper cell edges and improves the signal‑to‑noise ratio for downstream classification.

2. Multi‑wavelength DHM – Holograms are recorded at three distinct wavelengths (≈ 450 nm, 532 nm, 650 nm). By jointly reconstructing the complex fields and exploiting the wavelength‑dependent scattering, contrast between nuclear material, granules, and cytoplasm is enhanced, which in turn raises the CNN’s feature discrimination power.

3. Physics‑Informed Neural Network (PINN) reconstruction – A PINN is trained to satisfy the Helmholtz wave equation while minimizing the difference between the predicted and measured hologram intensity. Early experiments show reduced speckle and more faithful phase recovery compared with standard FFT‑based propagation, suggesting a path toward real‑time, low‑power reconstruction on embedded hardware.

End‑to‑end workflow and clinical validation
A complete pipeline is demonstrated on real patient blood samples: (i) minimal sample preparation (dilution and loading into a disposable cartridge), (ii) hologram acquisition (≈ 100 ms exposure), (iii) automated cell segmentation using a watershed‑based algorithm, (iv) CNN classification, and (v) MDW calculation. The total turnaround time per sample is 2–3 minutes, fitting within the workflow of existing blood‑gas analyzers. The study reports a modest dataset (≈ 1,200 annotated cells) and acknowledges that larger, multi‑center trials are required to confirm robustness across diverse pathologies and demographic groups.

Limitations and future directions
Key challenges identified include: (a) sensitivity of the lens‑less design to laser speckle and sensor noise, (b) limited representation of rare cell types (e.g., platelets, nucleated red blood cells) in the training set, (c) computational load of PINN reconstruction for real‑time operation, and (d) integration of automated fluidics for fully hands‑free sample handling. The authors propose scaling up the dataset, exploring speckle‑reduction illumination schemes, implementing model compression (e.g., quantization, knowledge distillation) for edge deployment, and developing a custom ASIC or FPGA accelerator for PINN inference.

Overall impact
The dissertation convincingly demonstrates that DHM, when coupled with modern deep‑learning classifiers, can replace traditional bright‑field microscopy for differential white‑blood‑cell counting in a compact, potentially disposable PoC device. The lens‑less architecture, together with the ability to extract MDW, positions the technology as a dual‑function analyzer (blood‑gas + hematology) that could dramatically shorten diagnostic latency in intensive care units, operating rooms, and resource‑limited settings. By also delivering a suite of ancillary optical innovations (super‑resolution, multi‑wavelength contrast, physics‑aware reconstruction), the work establishes a versatile platform that can be extended to other label‑free cytology applications, such as malaria detection or circulating tumor cell enumeration. Continued development along the outlined pathways could lead to regulatory‑cleared, commercially viable products within the next 3–5 years.