NEUROMORPHIC-ENABLED VIDEO-ACTIVATED CELL SORTING FOR HIGH-ACCURACY CLASSIFICATION OF REGULAR RED BLOOD CELLS AND BLOOD-DISEASE-RELATED SPHEROCYTES

Weihua He, Junwen Zhu, Yongxiang Feng, Fei Liang, Wenhui Wang2024 · DOI

Abstract

The image-activated cell sorter (IACS), serving as a comprehensive setup of imaging flow cytometry (IFC), holds extensive utility for cellular heterogeneity investigation. However, current IACS frameworks encounter challenges like 3D information loss and processing latency during real-time sorting. Previously, we established the neuromorphic-enabled video-activated cell sorter (NEVACS) in a reducing-redundancy-for-efficiency strategy and applied in classifying Hela cells and beads. Herein, we further implemented NEVACS by using spiking neural network (SNN) models and deployed them on a brain-inspiring neuromorphic chip to replace the previous implementation of GPU. With this new implementation, we applied NEVACS in the application of accurately sorting regular red blood cells (RBCs) and blood-disease-relevant spherocytes, highlighting the NEVACS' superior accuracy over IACS.

Introduction

In medical diagnostics, Imaging Flow Cytometry (IFC) is the gold standard for analyzing cell populations. However, traditional Image-Activated Cell Sorters (IACS) typically rely on a single 2D snapshot of a cell as it passes through a microfluidic channel. For non-spherical cells like Red Blood Cells (RBCs)—which are shaped like biconcave disks—this leads to significant "3D information loss." Depending on the cell's orientation, a healthy RBC might be misidentified as a spherocyte (a spherical cell associated with disorders like hereditary spherocytosis). Furthermore, processing high-resolution images in real-time usually requires power-hungry GPUs, creating latency and limiting portability.

Methods

The researchers developed an updated implementation of NEVACS (Neuromorphic-Enabled Video-Activated Cell Sorter) to address these limitations. The system architecture integrates several novel components:

  1. Event-Based Imaging: Instead of standard cameras, the system uses event cameras that only record changes in pixel intensity, drastically reducing data redundancy.
  2. Controlled Cell Rolling: The microfluidic chip employs a "vertical sheath fluid" to force cells to roll, ensuring the camera captures multiple angles and recovers 3D morphological data.
  3. Neuromorphic Computing: The system uses the HP201 neuromorphic chip, a 60-core architecture designed to run Spiking Neural Networks (SNNs). SNNs mimic biological neurons, providing extreme efficiency for processing temporal video data.
SNN architecture mapping on the HP201 chip and performance comparison against a standard GPU.

Results and Discussion

The system was evaluated using mixtures of regular RBCs and spherocytes at various concentrations (0%, 5%, 15%, 50%, and 80%).

Superior Diagnostic Accuracy

NEVACS demonstrated significantly higher accuracy than traditional IACS methods. In a critical 5% spherocyte test—a threshold vital for clinical diagnosis—traditional snapshot-based methods reported a massive overestimation of 28.56%. In contrast, NEVACS achieved a near-perfect result of 4.23%.

Comparison of spherocyte detection accuracy between NEVACS and traditional IACS methods across different sample compositions.

Efficiency and Throughput

By moving from a GPU (Nvidia RTX3080) to the HP201 neuromorphic chip, the system achieved substantial performance gains:

  • Latency: Inference time dropped from ~676 µs (GPU) to ~395 µs (HP201).
  • Throughput: The system reached a total throughput of approximately 1,000 cells per second.
  • Power Consumption: The neuromorphic setup consumed only 14.36 W, compared to the 137.55 W required by the GPU—a nearly 10x reduction.

Conclusion

NEVACS successfully leverages hardware-software synergy to overcome the limitations of traditional cell sorting. By intentionally inducing cell rolling and processing the resulting temporal signatures via SNNs, the system achieves high-accuracy classification with minimal power requirements. Future work will focus on scaling the system for ultra-high-throughput industrial applications and expanding the SNN architectures to handle more complex, multi-type biological samples. The low power profile makes NEVACS particularly suitable for future portable, point-of-care diagnostic devices.

Summary updated

Abstract

The image-activated cell sorter (IACS), serving as a comprehensive setup of imaging flow cytometry (IFC), holds extensive utility for cellular heterogeneity investigation [1].However, current IACS frameworks encounter challenges like 3D information loss and processing latency during real-time sorting.Previously [2], we established the neuromorphic-enabled video-activated cell sorter (NEVACS) in a reducing-redundancy-for-efficiency strategy and applied in classifying Hela cells and beads.Herein, we further implemented NEVACS by using spiking neural network (SNN) models and deployed them on a brain-inspiring neuromorphic chip to replace the previous implementation of GPU.With this new implementation, we applied NEVACS in the application of accurately sorting regular red blood cells (RBCs) and blood-disease-relevant spherocytes, highlighting the NEVACS' superior accuracy over IACS.
BibTeX
@misc{he2024neuromorphic,
  author = {Weihua He and Junwen Zhu and Yongxiang Feng and Fei Liang and Wenhui Wang},
  title = {{NEUROMORPHIC-ENABLED VIDEO-ACTIVATED CELL SORTING FOR HIGH-ACCURACY CLASSIFICATION OF REGULAR RED BLOOD CELLS AND BLOOD-DISEASE-RELATED SPHEROCYTES}},
  year = {2024},
  doi = {10.70477/shex9590},
}
Figure