經顱光生物調節光通量於前額葉皮質的非侵入性定量

Abstract

本研究的主要目標是透過近紅外光譜技術獲得漫反射光譜，以非侵入式的方式量化經顱光生物調節過程中前額葉皮質的光通量密度。基於提出的神經網路預測模型，使用頭皮表面的空間解析漫反射光譜，結合人口統計參數和頭部物理測量數據作為特徵，來預測刺激光束穿透前額葉皮質的光子能量比例。本研究提供了一種最佳化個人劑量的非侵入性替代方案，相較於傳統的固定劑量，使用神經網路模型可以有效減少與真實值之間的誤差。

在訓練模型的方法上，本研究使用 154 名年齡介於 51 到 80 歲間的健康受試者的 MRI 頭部掃描影像建立 3D 模型，根據頭皮、顱骨、腦脊髓液、灰質和白質等主要組織層的均勻光學參數變化，並使用多波長和不同偵測器與光源間距，以蒙地卡羅模擬漫反射光譜，作為訓練資料。神經網路架構結合 SLSTM 模型以及全連接網路，用於預測光源波長在 810 nm 和 1064 nm 時，目標關注區域內的平均光通量密度。

本研究所建立的預測模型結果顯示，不同預測驗證實驗設計下的相對誤差表現良好，特別是在 New-OPs 實驗中，使用所有生成資料中不同組別的吸收和散射係數組合分別作為訓練集和測試集，達到了最低的相對誤差水平。此外，在接近真實情境的受試者交叉驗證實驗中，由於受試者樣本數量有限，相對誤差較高，反映了模型在捕捉人類頭部解剖結構多樣性方面的挑戰。然而，與固定刺激劑量方法相比，整體上顯著提升了對不同受試者光通量密度的預估準確性，平均使相對誤差降低約 27% 至 49%。尤其針對在固定劑量方法中相對誤差較大的受試者，預測結果明顯降低相對誤差，達到約 94% 至 112% 的改善。藉由本研究提出的預測模型能有效減少過量和劑量不足的情況，為經顱光生物調節在臨床應用中的個人化治療提供了重要的依據。

The main objective of this study is to obtain diffuse reflectance spectra through near-infrared spectroscopy to non-invasively predict the photon fluence rate in the prefrontal cortex during transcranial photobiomodulation (tPBM). Based on the proposed neural network prediction model, spatially resolved diffuse reflectance spectroscopy from the scalp surface is used. These data, combined with demographic parameters and physical measurements of the head, serve as features to predict the proportion of photon energy penetrating the prefrontal cortex. This study provides a non-invasive alternative to optimize individual dosages, effectively reducing the error between the predicted and actual values compared to conventional constant dosages.

In the methodology, the study utilized MRI head scans of 154 healthy subjects, aged between 51 and 80 years, to build 3D models. These models simulate the absorption and scattering characteristics of major tissue layers, including the scalp, skull, cerebrospinal fluid, gray matter, and white matter, using Monte Carlo simulations to generate diffuse reflectance spectra as the training database. The simulations covered these tissues' homogeneous optical parameter variations, employing multiple wavelengths and different source-detector separations. The neural network architecture integrates an SLSTM model, along with fully connected networks, to predict the average photon fluence rate in the target regions at wavelengths of 810 nm and 1064 nm.

The results of the prediction model developed in this study showed good relative error performance under various validation experiments. This was especially evident in the New-OPs experiment, which achieved the lowest relative error. In this experiment, different combinations of absorption and scattering coefficients were used in the generated database for the training and test sets, respectively. Additionally, in subject-wise cross-validation experiments that mimic real-life scenarios, the relative error was higher due to the limited number of subject samples. This reflects the challenges of the model in capturing the diversity of human head anatomical structures. However, compared to the constant stimulation dosage method, the overall accuracy of predicting photon fluence rates for different subjects significantly improved, reducing the relative error by approximately 27% to 49% on average. For subjects with higher relative errors in the fixed-dose method, the prediction results significantly reduce the relative error, achieving improvements of about 94% to 112%. The proposed predicted model effectively reduced instances of over-dosing and under-dosing. This provides important support for personalized treatment in clinical applications of tPBM.