利用電腦模擬磁振頻譜與可解釋特徵之卷積神經網路預測腦中代謝物之濃度

Abstract

磁振頻譜可用於量測大腦中代謝物之濃度，然而定量代謝物之濃度需要取得詳細完整的信號，並經過一連串的信號處理，誤差可能會從中不斷累積。近年來已有許多研究將深度學習應用在信號強度之預測。本研究使用了卷積神經網路搭配電腦模擬生成之頻譜進行腦中代謝物之絕對定量，包含五種常見代謝物: Cr、 tCho、tNAA、Glu、mI，並且與常用之定量軟體LCMODEL比較定量結果。一般很難解釋深度學習模型所學到的事物，本研究使用saliency map 與filter activation嘗試將CNN模型學習到的特徵視覺化。在電腦模擬生成之頻譜中，上述五種代謝物之理想信號先以線性組合疊加，再增加譜線寬度、加上複雜起伏的基線以及不同程度的雜訊與相位誤差模擬實際信號的缺陷，每個頻譜會搭配一個未壓抑之水訊號以計算代謝物之絕對濃度。本研究預測了模擬頻譜以及真實人類大腦的頻譜中代謝物之絕對濃度，並以五種代謝物之總均方根誤差(RMSE)以及每一種代謝物之平均絕對百分誤差 (MAPE)、RMSE、皮爾森相關係數與t檢定評估模型之預測準確度。Saliency map透過將損失函數對影像中的每一個像素做偏微分，可以看見模型在做預測時，影像中的哪些地方會有顯著的影響。而filter activation視覺化了模型中每層卷積層中filter的輸出。若只使用模擬頻譜訓練模型，定量真實頻譜之平均總RMSE為32%左右。然而在將占總訓練資料數約2%之真實data加入訓練資料後，平均總RMSE明顯降低至23%左右。當使用模擬頻譜作為訓練資料的數目在1000~10000時，可達到較準確的定量結果，模擬頻譜過多或過少均會使準確度下降。Saliency map的結果顯示使用模擬頻譜能幫助模型找出所預測代謝物之位置，否則在mI這種信雜比較低之代謝物中模型的預測可能會被其他信號「誤導」。在filter activation的結果中，強度較高的訊號較容易被保留下來，但是強度較高且非代謝物訊號的假影也可能被保留下來。本研究提出了一個磁振頻譜定量之模型，可預測五種主要代謝物之絕對濃度且準確度能與傳統方法比較，並提出了簡化方法，最後將模型視覺化解釋其所學習之特徵。

MRS is a technique for in-vivo measurement of metabolite concentrations. Quantification of metabolite concentrations from magnetic resonance spectroscopy requires detailed signal formulae and a series of signal processing, through which error may propagate. Recently, convolutional neural network (CNN) has been explored for its ability to estimate signal intensity/contrast without explicitly specified biophysical models. The goal of this study is to estimate the absolute metabolite concentrations of the brain by using synthetic data and convolutional neural network (CNN), then provide comparison with LCMODEL. However, it remains unclear how the network learns from the data. This study aimed to understand the pick-up of features by using the gradient-based saliency map and filter activations. In this study, five metabolites were included: creatine (Cre), glutamate (Glu), myo-inositol (mI), total choline (tCho), and total N-acetylaspartate (tNAA). In synthetic spectra, the ideal spectra of the abovementioned five metabolites were linearly combined and then corrupted by line broadening, complex baseline, and varied levels of noise and phase shift. The performances of the models were evaluated by total root mean square percentage error (RMSE) of five metabolites, RMSE, mean absolute percent error (MAPE) and Pearson’s correlation coefficient for each metabolite. A student’s t-test was performed. For quantification of in-vivo spectra, data show that if only simulated spectra were used for training, the total RMSE between the labeled and predicted concentrations is ~32%. However, with a small number of in-vivo spectra in the training set (~2%), the total RMSE is ~23% over five cerebral metabolites, which is notably lower. In this study, the ideal range for number of simulated spectra in training set is 1000~10000. The quantification error increased if the number of simulated spectra in training set was too little or too much. The saliency map further indicated that synthetic data helped directing the model to “look at” the metabolite whose concentration was to be predicted. The improvement was most noticeable for mI which the model seemed to ignore likely due to the distraction from Cho and/or Cr. In conclusion, we proposed a CNN model, which is able to predict five cerebral metabolites that are commonly investigated in neurological diseases with an overall performance comparable with that of conventional model-based fitting. We demonstrated the visualizations of CNN models by saliency maps and filter activations to understand the pick-up of features from the in-vivo data.