Fast Analysis of Time-Domain Fluorescence Lifetime Imaging via Extreme Learning Machine

Zhenya Zang, Dong Xiao, Quan Wang, Zinuo Li, Wujun Xie, Yu Chen, David Day-Uei Li2022 · DOI

TL;DR

This paper introduces an Extreme Learning Machine (ELM) approach for significantly accelerating the analysis of fluorescence lifetime imaging microscopy (FLIM) data. By leveraging ELM's efficient learning capabilities, the method achieves comparable accuracy to existing algorithms with substantially reduced training and inference times, making it promising for real-time applications and embedded systems.

What Problem Does This Paper Address?

Fluorescence Lifetime Imaging Microscopy (FLIM) is a powerful technique for visualizing biological processes by measuring how long fluorophores emit light after excitation. However, extracting accurate lifetime information from FLIM data is computationally intensive, often requiring lengthy acquisition and processing times. This limits its application in dynamic biological studies and real-time diagnostics. Existing methods, including iterative fitting and even some neural network approaches, struggle with speed, computational complexity, and robustness under low-photon conditions, which are common in sensitive biological imaging to avoid photobleaching. The paper aims to bridge this gap by developing a faster, more accurate, and robust FLIM analysis method.

Methods

The core of this work is the application of the Extreme Learning Machine (ELM) algorithm for FLIM data analysis. ELM is a type of single-hidden-layer feed-forward neural network that differs from traditional neural networks in its training approach. Instead of using back-propagation, ELM randomly assigns weights to the hidden layer and then analytically determines the output weights using a Moore-Penrose matrix inversion. This "data-driven" approach without back-propagation significantly speeds up training and inference.

The authors adapted ELM to process 1D histograms from Time-Correlated Single-Photon Counting (TCSPC) measurements, which are standard in FLIM. They specifically focused on analyzing mono- and bi-exponential decay models, common in biological samples. The ELM model takes the photon arrival histogram as input and outputs the fluorescence lifetime parameters (e.g., τ₁, τ₂, and amplitude α).

Datasets

The study utilized two types of datasets:

  • Synthetic Datasets: 20,000 synthetic histograms were generated to simulate FLIM data. These datasets were designed to mimic real TCSPC measurements, including noise and instrument response functions. They were used to rigorously compare ELM against other algorithms under controlled conditions, including varying photon counts.
  • Experimental Datasets: Living prostate cancer cells incubated with functionalized gold nanorods (GNRs) labeled with Cy5 were used. This provided a realistic biological sample for evaluating ELM's performance in a complex environment.

Key Insights and Ideas

  • ELM's Speed Advantage: The most striking insight is ELM's dramatically reduced training and inference times compared to traditional fitting methods (like NLSF) and even other neural network architectures (like 1D-CNN). This is attributed to its non-back-propagation training strategy.
  • Robustness in Low-Photon Conditions: ELM demonstrated remarkable performance even with very low photon counts, a critical factor for imaging delicate biological samples without causing photobleaching. It could still reconstruct meaningful lifetime images where intensity images lost crucial details.
  • Accuracy Comparable to Advanced Methods: Despite its speed, ELM achieved accuracy comparable to, and often superior to, established methods like NLSF and even a 1D-CNN, particularly in challenging scenarios.
  • Flexibility and Simplicity: ELM's architecture is simpler than many deep learning models, requiring less hyperparameter tuning. This makes it more adaptable and easier to implement, especially for embedded systems.
  • Potential for Online Training: The fast training speed of ELM opens up possibilities for online training directly on embedded hardware, enabling real-time adaptation and analysis.

Key Results

  • Synthetic Data: ELM consistently outperformed NLSF in terms of absolute error and F-value across various photon count levels and for both mono- and bi-exponential decays. ELM also showed lower bias compared to NLSF, especially at lower photon counts.
  • Inference Time: For bi-exponential decay analysis, ELM took only 6.5 seconds, while NLSF took 670.9 seconds. Even compared to a 1D-CNN running on a GPU, ELM on a CPU was significantly faster (1.73 s vs. 116.43 s).
  • Experimental Data: ELM successfully reconstructed lifetime images of prostate cancer cells and gold nanoprobes, showing good agreement with a 1D-CNN and outperforming NLSF, which exhibited significant bias. BCMM, while fast, was found to be less robust to varying photon counts.
  • Low-Photon Scenario: In experiments with reduced scanning cycles (down to 10), ELM was able to reconstruct lifetime images that revealed cell boundaries and nanoprobes, information lost in the corresponding intensity images.

Limitations and Future Work

  • Model Complexity: While ELM can handle mono- and bi-exponential decays, the paper primarily focuses on these. Extending it to more complex multi-exponential decays might require further architectural adjustments.
  • Hardware Implementation: While the paper suggests ELM is promising for embedded hardware, the actual implementation and performance on specific embedded platforms are not detailed.
  • Future Work: The authors highlight the potential for online training and implementation on embedded hardware, suggesting future work in this direction. They also mention the promise of ELM for applications like cellular cancer diagnosis, implying further biological validation and application development.

Summary updated

BibTeX
@misc{zang2022fast,
  author = {Zhenya Zang and Dong Xiao and Quan Wang and Zinuo Li and Wujun Xie and Yu Chen and David Day-Uei Li},
  title = {{Fast Analysis of Time-Domain Fluorescence Lifetime Imaging via Extreme Learning Machine}},
  year = {2022},
  doi = {10.3390/s22103758},
}
Figure