用於生成體內非相干運動成像的無監督一維卷積神經網路、雙向長短期記憶網路和多頭注意力模型

Abstract

體內非相干運動 (IVIM) 成像是一種先進的擴散磁振造影 (MRI) 技術，廣泛應用於定量分析組織微結構中的灌注與擴散情況。然而，傳統的逐像素擬合方法在計算過程中往往會因為受到高雜訊的影響而導致參數的估計結果不佳。為解決此問題，本研究提出了一種結合了一維卷積神經網絡、雙向長短時記憶網絡和多頭注意力機制的無監督學習框架。該模型在模擬和真實擴散加權 MRI 資料上進行了評估，並將其結果與使用信任區域反射 (TRR)、深度神經網路 (DNN) 和貝葉斯-瑪律科夫隨機場 (Bayesian-MRF) 方法獲得的結果進行了比較。模擬結果表明，在大多數情況下，所提出的方法比其他三種方法能達到更可靠的參數估計。即使在高信噪比 (即 100) 的條件下，所提出的方法仍能達到比 TRR 方法和DNN方法更低的均方根誤差值，且其結果與Bayesian-MRF方法的結果相當。在真實腹部資料集實驗中，相較於其他三種方法，所提出的方法得出的IVIM參數估計值也接近文獻報導的值。對於真實腦部資料集，建議的方法和 Bayesian-MRF方法的表現都優於其他方法，並且前者在所有IVIM參數上都達到最佳的對比訊噪比。這些研究結果表明，我們所提出的方法是一種很有前途的IVIM參數估計方法。

Intravoxel incoherent motion (IVIM) imaging, an advanced diffusion magnetic resonance imaging (MRI) technique, is widely used to quantify perfusion and diffusion in tissue microstructure. However, the traditional pixel-by-pixel fitting method is easily affected by high-level noise, leading to unreliable parameter estimation. To address this limitation, an unsupervised learning framework combining with a one-dimensional convolutional neural network, a bidirectional long short-term memory network, and a multi-head attention mechanism was proposed. The model was evaluated on both simulated and real diffusion-weighted MRI data. The results were compared with those obtained using the trust region reflective (TRR) method, the unsupervised deep neural network (DNN), and the Bayesian-Markov random fields (Bayesian-MRF) method. Simulation results indicate that the proposed method could provide more reliable parameter estimation than the other methods under most cases. Even at high signal-to-noise ratios levels (i.e., 100), the proposed method achieved lower root mean square error values than the TRR method and the DNN method, and performed comparably to the Bayesian-MRF method. For the abdominal datasets, the proposed method yielded IVIM parameter estimates generally closer to the literature reported values compared to those derived from the other three methods. For the brain dataset, both the proposed method and the Bayesian-MRF method outperformed the other methods, with the former achieving better contrast-to-noise ratios for all IVIM parameters. These findings suggest that our proposed method is a promising approach for IVIM parameter estimation.