Cerebral blood flow monitoring using a deep learning implementation of the two-layer diffuse correlation spectroscopy analytical model with a 512 × 512 SPAD array

Tools

Pan, Mingliang and Li, Chenxu and Zhang, Yuanzhe and Mollins, Alan and Wang, Quan and Erdogan, Ahmet T. and Hua, Yuanyuan and Zang, Zhenya and Finlayson, Neil and Henderson, Robert K. and Li, David Day-Uei (2025) Cerebral blood flow monitoring using a deep learning implementation of the two-layer diffuse correlation spectroscopy analytical model with a 512 × 512 SPAD array. Neurophotonics, 12 (3). 035008. (https://doi.org/10.1117/1.NPh.12.3.035008)

[thumbnail of Pan-etal-Neurophotonics-2025-Cerebral-blood-flow-monitoring-using-a-deep-learning-implementation]
Preview
 Text. Filename: Pan-etal-Neurophotonics-2025-Cerebral-blood-flow-monitoring-using-a-deep-learning-implementation.pdf
Final Published Version
License: Creative Commons Attribution 4.0 logo
Download (12MB)| Preview

Abstract

Significance: Multi-layer (two- and three-layer) diffuse correlation spectroscopy (DCS) models improve cerebral blood flow index (CBFi) measurement sensitivity and mitigate interference from extracerebral tissues. However, their reliance on multiple predefined parameters (e.g., layer thickness, optical properties) and high computational load limits their feasibility for real-time bedside monitoring. Aim: To develop a fast, accurate DCS data processing method based on the two-layer DCS analytical model, enabling real-time cerebral perfusion monitoring with enhanced brain sensitivity. Approach: We employed deep learning (DL) to accelerate DCS data processing. Unlike previous DCS networks trained on single-layer models, our network learns from the two-layer DCS analytical model, capturing extracerebral vs. cerebral dynamics. Realistic noise was estimated from subject-specific baseline measurements using a 512×512 SPAD array at a large source-detector separation (35 mm). The model was evaluated on test datasets simulated with a four-layer slab head model via Monte Carlo (MC) methods and compared against conventional single-exponential fitting and the two-layer analytical fitting. Two in-vivo physiological response tests were also conducted to assess the real-world performance. Results: The proposed method bypasses traditional curve-fitting, achieved real-time monitoring of CBF changes at 35 mm separation for the first time with a DL approach. Validation on MC simulations shows superior accuracy in relative CBFi estimation (4.1% error vs. 12.7% for single-exponential fitting) and significantly enhanced CBFi sensitivity (86.5% vs. 57.7%). Although the two-layer analytical fitting offers optimal performance, it depends on strict assumptions and preconditions, and its computational complexity makes it time-consuming and unsuitable for real-time monitoring applications. Additionally, our method minimizes the influence of superficial blood flow and 750-fold faster than single-exponential fitting in a realistic scenario. In-vivo tests further validated the method’s ability to support real-time cerebral perfusion monitoring and pulsatile waveform recovery. Conclusions: This study demonstrates that integrating DL with the two-layer DCS analytical model enables accurate, real-time cerebral perfusion monitoring without sacrificing depth sensitivity. The proposed method enhances CBFi sensitivity and recovery accuracy, supporting future deployment in bedside neuro-monitoring applications.

ORCID iDs

Pan, Mingliang ORCID logoORCID: https://orcid.org/0009-0001-9732-8963, Li, Chenxu, Zhang, Yuanzhe ORCID logoORCID: https://orcid.org/0009-0004-5107-3856, Mollins, Alan ORCID logoORCID: https://orcid.org/0009-0002-2328-1207, Wang, Quan ORCID logoORCID: https://orcid.org/0009-0005-2936-825X, Erdogan, Ahmet T., Hua, Yuanyuan, Zang, Zhenya, Finlayson, Neil, Henderson, Robert K. and Li, David Day-Uei ORCID logoORCID: https://orcid.org/0000-0002-6401-4263;
CORE (COnnecting REpositories)