擴散相關光譜術定量人體腦血流指數之模擬研究

Abstract

擴散相關光譜術（Diffuse Correlation Spectroscopy, DCS）是一種利用近紅外光量測血管血流變化的非侵入性技術，具有可攜式、可連續監測等優點。然而，傳統連續波DCS（Continuous Wave DCS, CW-DCS）對於深層腦組織量測靈敏度有限，且量測結果容易受到頭皮與顱骨等淺層組織及光學參數不確定性的影響，使得腦血流指數（Blood Flow Index, BFi）的定量仍具挑戰性。本研究旨在建立一套適用於人體頭部的CW-DCS模擬與分析流程，並透過神經網路加速順向模型的計算，系統性評估光學參數誤差對BFi估測之影響，進一步量化BFi之不確定性，藉以提升腦血流定量結果的可靠性。 本研究首先使用磁振造影（Magnetic Resonance Imaging, MRI）建立含頭皮、顱骨、腦脊髓液、灰質與白質之多層頭部模型，使用蒙地卡羅法模擬不同光源與偵測距離（Source-Detector Separation, SDS）下的自相關訊號，評估模擬穩定性與統計誤差。接著以模擬資料訓練類神經網路（Artificial Neural Network, ANN）順向模型，學習由光學參數與BFi預測強度自相關函數，並搭配非線性曲線擬合，擬合頭皮和灰質層組織之BFi，以建立絕對血流指數(BFi)的估測流程。而相對腦血流指數(ΔBFi)預測模型的部分則是透過特徵萃取與機器學習方法，預測BFi的相對變化量。 在誤差分析方面，本研究針對吸收係數及散射係數等光學參數施加 ±20% 的誤差，評估不同層次組織對BFi估測的敏感度。結果顯示，頭皮和灰質之光學參數誤差會放大傳遞至BFi的估計，散射係數的誤差更可能造成40%以上的BFi偏差，顯示準確量測或校正光學參數為腦血流定量的關鍵。 綜合而言，本研究建立了一套以多層頭部模型、蒙地卡羅模擬與ANN順向模型為核心的CW-DCS分析架構，不僅量化了光學參數與頭部結構誤差對腦血流指數的影響，也驗證了以類神經網路加速自相關函數預測與BFi擬合的可行性。此結果可為未來DCS研究上提供設計與參數選擇上的參考。

Diffuse correlation spectroscopy (DCS) is a noninvasive technique that uses near-infrared light to measure changes in microvascular blood flow, offering advantages such as portability and the capability for continuous monitoring. However, conventional continuous-wave DCS (CW-DCS) has limited sensitivity to deep brain tissues, and its measurements are easily affected by superficial layers such as the scalp and skull, as well as uncertainties in tissue optical properties. These factors make quantitative estimation of the blood flow index (BFi) challenging. This study aims to establish a CW-DCS simulation and analysis framework for the human head and to employ neural networks to accelerate the forward-model computation, thereby enabling a systematic evaluation of the influence of optical-parameter errors on BFi estimation and subsequent quantification of BFi uncertainty, with the ultimate goal of improving the reliability of cerebral blood flow quantification. Magnetic resonance imaging (MRI) was first used to construct a multilayer head model comprising scalp, skull, cerebrospinal fluid, gray matter, and white matter. Monte Carlo simulations were then performed to generate intensity autocorrelation signals at various source–detector separations (SDS), and the stability of the simulations as well as the associated statistical errors were evaluated. On this basis, an artificial neural network (ANN)–based forward model was trained using the simulated data to learn the mapping from optical properties and BFi to the intensity autocorrelation function. The ANN forward model was further integrated with nonlinear curve fitting to retrieve the BFi of the scalp and gray-matter layers, thereby establishing a procedure for absolute BFi estimation. In addition, a relative cerebral blood flow index (ΔBFi) prediction model was developed using feature-extraction and machine-learning methods to estimate relative changes in BFi. For the error analysis, ±20% error were applied to optical parameters including the absorption coefficient, and scattering coefficient, in order to assess the sensitivity of BFi estimation to different tissue layers. The results show that errors in the optical properties of the scalp and gray-matter layers are amplified when propagated to the estimated BFi, and that errors in the scattering coefficient can lead to more than 40% bias in BFi. These findings indicate that accurate measurement or calibration of optical parameters is critical for quantitative assessment of cerebral blood flow. In summary, this research established a CW-DCS analysis framework centered on a multi-layer head model, Monte Carlo simulations, and an ANN forward model. This framework not only quantifies the impact of errors in optical parameters and head-structure modeling on the cerebral blood flow index, but also verifies the feasibility of using neural networks to accelerate autocorrelation prediction and BFi fitting. The results can provide guidance for experimental design and parameter selection in future DCS studies.