Compact and robust deep learning architecture for fluorescence lifetime imaging and FPGA implementation.

Area of Science:

• Biomedical Optics
• Deep Learning
• Computational Imaging

Background:

• Time-domain fluorescence lifetime imaging (FLIM) is crucial for biomedical diagnostics.
• Conventional FLIM methods often face challenges with computational complexity and data processing efficiency.

Purpose of the Study:

• To develop a computationally efficient deep learning network for FLIM.
• To improve the speed and accuracy of fluorescence lifetime retrieval.
• To explore hardware implementation for enhanced performance.

Main Methods:

• Proposed a 1D Fluorescence Lifetime AdderNet (FLAN) utilizing l1-norm extraction and avoiding multiplication-based convolutions.
• Introduced a log-scale merging technique (FLAN+LS) for temporal compression of fluorescence decays.
• Evaluated FLAN and FLAN+LS using synthetic and real confocal microscopy data.
• Implemented the network on a field-programmable gate array (FPGA) with post-quantization.

Main Results:

• FLAN+LS achieved compression ratios of 0.11 and 0.23 compared to FLAN and 1D CNN, respectively, with high lifetime retrieval accuracy.
• Networks demonstrated minor reconstruction errors across various photon-count scenarios.
• Hardware implementation on FPGA showed superior computing efficiency for FLAN+LS over 1D CNN and FLAN.
• Effectiveness validated by differentiating fluorescent beads with distinct lifetimes.

Conclusions:

• The proposed FLAN and FLAN+LS networks offer a computationally efficient and accurate solution for FLIM.
• Hardware acceleration via FPGA significantly enhances the practical applicability of these deep learning models.
• The developed architecture shows promise for other time-resolved biomedical sensing applications.