代理模型加速蒙地卡羅模擬及類神經網路定量內頸靜脈血氧變化量

Abstract

本研究的主要目標在於透過近紅外光譜量測技術，以非侵入式的方式定量人體內頸靜脈血氧飽和度變化量，首先透過基於類神經網路的代理模型對傳統組織光學模擬所使用的蒙地卡羅法進行加速，接著應用類神經網路建立血氧飽和度變化量預測模型，其輸入為經由公式萃取出光譜特徵的特徵光譜，輸出為血氧飽和度變化量。 在實體量測系統上，本研究使用20個波長點根據血液的吸收光譜特徵設立，波長範圍介在700nm~850nm之間，並且建立雙通道系統，短通道的部分，光源與偵測器的距離為10mm，長通道的部分，光源與偵測器的距離為20mm，透過這樣的設計能夠有效降地淺層組織的影響並且放大深層組織及內頸靜脈所在的區域的訊號。在模擬測試時是基於受試者的頸部超音波影像建立的三維數值模型，讓模擬結果能與現實更加貼近，得到更加準確的模擬資料。 本研究所建立的預測模型根據模擬結果，預測出內頸靜脈的血氧飽和度變化量其RMSE<1.5%。模型效能的評估上，本研究對人體呼吸造成內頸靜脈管徑大小與深度改變、周遭組織血氧變化、量測訊號產生的誤差等等，對於預測模型產生的影響進行了實驗與調查，結果顯示呼吸造成的影響最大可能造成3%~4%的方均根誤差(root-mean-square error, RMSE)的上升，而周遭組織血氧的變化對於預測模型的預測效能影響並不顯著，最多只會有1%的RMSE上升，而若是量測到的訊號受到誤差影響，則會造成1%~2%的RMSE上升。 模型泛用化上，本研究透過遷移學習的方式進行模擬實驗，經由實驗觀察發現，使用與原先資料相比僅佔千分之一的資料集透過遷移學習的方式能夠得到RMSE=3.5% 的結果，而若是不使用遷移學習單純使用千分之一的資料集則會得到RMSE=7%的結果。 在活體實驗上，將活體量測到的漫反射光譜，根據本實驗所設計的公式萃取出其光譜特徵後經由適當的標準化後，輸入至預測模型進行血氧飽和度變化量的預測，其預測結果與活體光譜觀察到的現象有一致性。

The primary objective of this study is to quantitatively measure changes in internal jugular vein oxygen saturation non-invasively using near-infrared spectroscopy. Initially, a surrogate model based on neural networks is employed to accelerate the Monte Carlo method which is traditionally used to simulate photon transport in tissue. Subsequently, another neural network is applied to establish a predictive model for oxygen saturation changes. The input to this model consists of spectral features extracted using formulas same as modified Beer-Lambert law, while the output represents oxygen saturation changes. As for the measurement system, the study utilizes 20 wavelength points based on the absorption spectra of blood, within the wavelength range of 700 nm to 850 nm. A dual channel system is set up, with the short channel having a distance of 10 mm between the light source and detector, and the long channel having a distance of 20 mm. This design effectively minimizes the impact of superficial tissues and enhances the signal from deeper tissues including the internal jugular vein area. During simulation, a three-dimensional numerical model is constructed based on ultrasound images of each subject’s neck, ensuring that simulation results closely resemble reality, thus providing more accurate simulated data. To evaluate the prediction model’s performance, the study investigates the impacts of factors such as human respiration, changes in oxygen levels in surrounding tissues, and measurement noise on the predictive model. The results indicate that the effects of respiration may lead to a maximum increase of 3% to 4% in root-mean-square error (RMSE). Changes in oxygen levels in surrounding tissues have a less significant impact, with a maximum RMSE increase of only 1%. Measurement signal errors can cause an RMSE increase of 1% to 2%. For model generalization, the study conducts simulated experiments using transfer learning. Through experimentation, it is observed that by using a thousandth of the original dataset and employing transfer learning, an RMSE of 3.5% can be achieved, while without transfer learning and using only a thousandth of the dataset, an RMSE of 7% is obtained. Based on the simulation results, the prediction model established in this study predicts changes in internal jugular vein oxygen saturation with an RMSE of less than 1.5%. In vivo experiments involve measuring diffuse reflectance spectra from living subjects, extracting spectral features using the formulas designed in this study, and inputting them into the prediction model after appropriate normalization. The prediction results are consistent with expected physiological response and spectral features in the measured data.