The development of the real-time algorithm in diffuse correlation spectroscopy for blood flow measurement

Abstract

Blood flow is fundamental for the distribution of oxygen and nutrients across the body, as well as the removal of metabolic waste from tissues. A sufficient microvascular blood supply is essential for proper tissue function. When the blood flow is obstructed due to vascular stenosis or other factors, cardiovascular and cerebrovascular diseases may develop. Continuous monitoring of blood flow is crucial both to prevent these diseases and to reduce associated healthcare costs. Blood flow measurement can be conducted through various imaging techniques (e.g., Positron Emission Tomography (PET), Arterial Spin Labeling Magnetic Resonance Imaging (ASL-MRI), Computed Tomography (CT)) and bio-measurement devices (e.g., Transcranial Doppler Ultrasonography (DU), Diffuse Correlation Spectroscopy (DCS)). While CT, PET, and MRI provide valuable three-dimensional blood flow information, they are limited in their capacity for continuous monitoring. Transcranial Doppler Ultrasonography is constrained in detecting changes at the microvascular level. DCS, by contrast, offers a non-invasive and ionizing-radiation-free method that allows for continuous bedside monitoring of blood flow in vivo, even in deep tissues. Its effectiveness has been validated through comparisons with other systems (e.g., ASL-MRI, Xe-CT, laser/ultrasound Doppler, contrast-agent bolus uptake). Within DCS methodologies, Continuous-Wave DCS (CW-DCS) employs a coherent laser source and single-mode or few-mode optical fibers in its detection system, using light scattering in a homodyne optical setup. This configuration provides a relatively simple means of measuring blood flow by analyzing scattered light from biological tissues. CW-DCS can quantify blood flow changes by constructing autocorrelation functions from the intensity of detected light and operating in a way that nonlinearly fits the photon diffusion equation. However, CW-DCS’s conventional blood flow index extraction method has limitation on the signal processing speed depending on the system, since the nonlinear fitting technique is used for this process. To overcome this drawbacks, new computational methods (e.g., the inverse tau method, software correlator, Field- programmable gate array based nonlinear fitting implementation, modified Beer-Lambert Law (MBLL) for blood flow) have been developed. The artificial intelligence models (e.g., convolutional neural network, long short-term memory-based models) have been also introduced for the new methods for reducing the signal processing time in CW-DCS. These methods increased computational speed but are vulnerable to noise, have limitation in the probing depth, and have not been widely tested across different blood flow conditions. CW-DCS also faces hardware-related limitations, particularly in terms of signal noise. As the distance between the light source and detector increases, the signal-to-noise ratio (SNR) decreases, reducing sensitivity and making measurements more susceptible to ambient light and patient movement—factors that limit its clinical utility. Alternative systems have been explored to address these hardware constraints, including interferometric detection, parallelized speckle detection, acousto-optic modulation, pathlength-resolved methods, speckle contrast techniques, and long-wavelength approaches. One promising method is long- wavelength interferometric DCS (LW-iDCS), which uses a high-speed line-scan camera with a 1064 nm wavelength source, multi-mode detection fiber, and free-space interferometer to improve SNR over CW-DCS. LW-iDCS requires a custom data analysis pipeline, including analog-to-digital converter offset subtraction, quadratic detrending, pixel averaging, and singular value decomposition (SVD) to remove raw data distortions caused by non-idealities in the line-scan camera, environmental vibrations, and motion. However, this analysis pipeline is time-consuming, limiting LW-iDCS’s ability to perform real-time blood flow measurements. To address these challenges, this study proposes real-time blood flow measurement methods to overcome the limitations of both CW-DCS and LW-iDCS systems. For CW-DCS, we introduce a numerical integration- based algorithm that maintains fast computation while addressing the drawbacks of conventional methods. Specifically, we propose the inverse of K2 (IK2), based on diffuse speckle contrast analysis, and the inverse of the numerical integration of squared g1 (INISg1), a simplified method that uses the normalized electric field autocorrelation curve. To reduce computation time further, g1 thresholding was applied, enabling these methods to match traditional nonlinear fitting approaches in performance. These methods were validated through simulations, liquid phantom studies, and in vivo experiments. Additionally, we tested these methods on three Arduino boards (Arduino Due, Arduino Nano 33 BLE Sense, and Portenta H7) to explore the potential for DCS system miniaturization through microcontroller-based signal processing. Results showed that both IK2 and INISg1 effectively captured blood flow changes, with INISg1 outperforming IK2 when g1 thresholding was applied. Implemented on both a PC and the advanced Portenta H7 board, INISg1 with g1 thresholding exhibited faster performance than current deep learning-based methods. Although this method requires signal gating and has yet to be tested across various noise levels, these findings suggest that INISg1 with g1 thresholding could provide a viable alternative for deriving relative blood flow information, contributing to DCS simplification. Furthermore, we developed a real-time blood flow measurement algorithm for the LW-iDCS system to replace the time-consuming data analysis pipeline. In LW-iDCS, SVD was previously used to remove independent components representing common signals across pixels, which might arise from reference intensity fluctuations or camera electronics and are unlikely to originate from the sample tissue. While SVD effectively reduces noise, it is computationally expensive and time-consuming, especially with large datasets. Moreover, tuning the number of components to be removed requires manual intervention, which is both labor- intensive and error-prone. To streamline blood flow estimation in LW-iDCS, we introduced a CNN model based on EfficientNet, a state-of-the-art architecture that optimizes accuracy and efficiency by uniformly scaling network depth, width, and resolution. Our model uses EfficientNet’s feature extraction capabilities to capture temporal information from raw data, which are transformed into three-channel images for EfficientNet’s input. The 1000 EfficientNet outputs are reduced to a single blood flow index. This model was trained and validated with blood flow indices from conventional LW-iDCS analysis, using data from breath- holding and tourniquet-induced pressure modulation maneuvers. The model’s results showed a high correlation (r² = 0.76, slope = 0.95) with conventional blood flow indices in both conditions. Although the model struggled to capture high-frequency components, it effectively eliminated the need for the computationally intensive SVD algorithm, thus simplifying data analysis. Additionally, it eliminated the need to calculate the autocorrelation function, further streamlining LW-iDCS analysis with high inference speed. Overall, this deep learning model offers a substantial improvement in LW-iDCS efficiency and usability, paving the way for real- time blood flow measurements. ©2025 Yoonho Oh ALL RIGHTS RESERVED