Towards high-performance deep learning architecture and hardware accelerator design for robust analysis in diffuse...

Area of Science:

• Biomedical Optics
• Computational Imaging
• Hardware Acceleration

Background:

• Diffuse correlation spectroscopy (DCS) is crucial for non-invasive blood flow monitoring.
• Accurate blood flow index (BFi) reconstruction is computationally intensive.
• Existing methods often require significant post-processing and lack miniaturization.

Purpose of the Study:

• To develop a compact deep learning (DL) architecture for BFi reconstruction in DCS.
• To create a highly parallelized computing hardware platform for real-time BFi calculation.
• To miniaturize DCS systems by integrating BFi reconstruction on-chip.

Main Methods:

• A lightweight DL architecture trained using simulated autocorrelation functions (ACFs) generated from an analytical model.
• Hardware implementation on FPGAs (Zynq-7000 and Zynq-UltraScale+) utilizing unrolling, pipelining, and pixel-wise parallelism.
• Simplified DL computing primitives using subtraction for feature extraction and fixed-point quantization.

Main Results:

• Achieved 66.7% and 18.5% improvement in mean squared error (MSE) for BFi and coherence factor β compared to CNNs on synthetic data.
• Enabled real-time processing of 10 and 15 ACFs per second on Zynq-7000 and Zynq-UltraScale+ FPGAs, respectively.
• Demonstrated an end-to-end on-chip conversion from intensity photon data to BFi and β, outperforming standalone hardware accelerators.

Conclusions:

• The proposed compact DL architecture and FPGA platform enable efficient, real-time BFi reconstruction for miniaturized DCS systems.
• This integrated on-chip solution eliminates the need for post-processing, enhancing the practicality of DCS.
• The computational efficiency of the FPGA accelerator was comprehensively compared against CPU and GPU solutions.