利用非線性迴歸模型估算多重B值之擴散權重影像

Estimation of Diffusion Weighted Image by Nonlinear Regression Model with Multiple B-Values

蔡明昇(Ming-Sheng Tsai)

指導教授 ： 陳泰賓

義守大學/電機資訊學院/資訊工程學系/碩士(2014年)

https://doi.org/10.6343/ISU.2014.00141

摘要

B-Value為擴散權重影像(Diffusion Weighted Image, DWI)中的梯度因子，是控制磁場的重要參數。對感興趣之區域給予最優化的參數條件，是建立B-value迴歸模型所追求的目標，利用後處理技術估算品質較佳的DWI影像，同時提供臨床較佳之診斷影像。
本研究旨在克服磁振造影掃描儀每次只能產生一組B值影像，若要產生N組B值或最佳B值之影像，必須進行N次DWI造影，耗時甚巨且代價高昂。研究使用GE Signa HD 1.5T 磁振造影儀(Magnetic Resonance Imaging, MRI)及八通道高解析度頭部線圈，並使用標準球形假體，材質為壓克力製成12cm x 12cm x12cm的球體，內含硫酸銅溶液(CuSO4)，進行全因子實驗設計收取影像資料；造影波序為EPI（Echo Planar Imaging），參數如下: 回音時間(Echo of Time, TE)為最小、可視範圍FOV為24 cm^2、切面厚度為5mm、切面間隔為1.5 mm、影像大小為256 x 256像素、激發次數(Number of Excitations, NEX)為1；實驗因子包括：9組B值、6組重複時間(Time of Repetition, TR)、是否採用平行造影技術(2組)。總計108組不同處理組合，並建立256 x 256 x 5 x 2個B值迴歸模型，當迴歸模型建立之後，再額外進行數個B值DWI影像之驗證實驗，其目的為比較迴歸估算之DWI影像進行差異性分析，其差異值再利用絕對相對誤差(Relative Absolute Error, RAE)進行結果判定，當RAE小於5%即可認定二者誤差很小，代表估算DWI(Estimation of DWI, DWIEst.)影像與真實DWI(True of DWI, DWITrue)影像差異小。
研究結果顯示，所發展之像素非線性迴歸模型技術，可順利虛擬欲求B值之擴散權重影像，測試其輸出的DWI與實際機器所取得的DWI之擬真度，RAE差異度僅6.43±1.17%，為使擬真度更加精準，加入AI人工智慧優勢判斷，進而將RAE差異度精進為5.47±0.43%，結果雖然未能將差異度控制在5%以內，但是AI技術確實大幅降低標準差，代表每組B值中5張影像之間沒有明顯差異。
建立像素非線性迴歸模型，進而估算其他B值之DWI並評估合適的B值，作為後續追蹤造影之參考。未來將改善人體各部位的影像品質及診斷精確度，對於學術研究及臨床醫學應用極具價值。

關鍵字

擴散權重影像 ； B值 ； 非線性迴歸模型 ； 絕對相對誤差 ； 人工智慧

並列摘要

B-Value is the gradient factor of Diffusion Weighted Images (DWI) and is also an important parameter to control the magnetic field. The goal of this study is to find optimized parameters to estimate of DWI by nonlinear regression model with multiple B-Values. Therefore, the good quality of DWI images is generated by presented techniques in order to provide clinical usefulness.
In this study, we try to conquer finding feasible B-value to generate high quality DWI images. In generally, conventional MRI have to scan N times in order to generate Nth DWI images. It is time consuming. The head coil of eight channel with the GE Signa HD 1.5T MRI was applied to standard spherical phantom which was made of 12cm x 12cm x 12cm acrylic fabric containing cupric sulfate solution (CuSO4). The pulse sequence is the Echo Planar Imaging (EPI). The used parameters are including, minimum echo time (TE), field of view (FOV) 24 cm2, thickness of slice 5 mm, spacing between slices 1.5 mm, image size 256 x 256 pixels, number of excitations (NEX) 1. The experimental factors include 9 B-Values, 6 time of repetition (TR), and whether using parallel imaging or not. Total of 108 different processing combinations and 256 x 256 x 5 x 2 regression models were employed in this study. Several additional DWI are imaging with B-values (DWITrue) in order to verify the estimated DWI by none linear regression models (DWIEst.). The relative absolute error (RAE) was used to exam the difference between DWITrue and DWIEst.. The RAE less than 5% means that DWIEst. is acceptable.
The phantom study shows that the presented approach can successfully estimate feasible B-values and simulated adaptive DWI. The average of RAE is about 6.43± 1.17%. The pixel based none linear regression model was built by artificial intelligence (AI) schema and resulting the average of RAE reduced to 5.47 ± 0.43%. The RAE provided by AI is not less than 5%, but AI technology does significantly reduce the standard deviation of estimated DWI. Hence, the presented method can estimate and simulate no significant diversity between five testing DWI images.
The pixel base nonlinear regression model not only estimate DWI images with small difference between testing DWI images, but find feasible B-value to simulate DWI images. In the future, the presented model might apply to real animal study in order to do verification.

並列關鍵字

DWI ； B-Value ； Nonlinear Regression Model ； RAE ； AI

參考文獻

2. Sakai K, Sakamoto R, Okada T, Sugimoto N, Togashi K. DWI based thermometry: the effects of b-values, resolutions, signal-to-noise ratio, and magnet strength. 34th Annual International Conference of the IEEE EMBS San Diego, California USA, 28 August - 1 September, 2012:2291-3.

3. Akazawa K, Yamada K, Matsushima S, Goto M, Yuen S, Nishimura T. Optimum b value for resolving crossing fibers: a study with standard clinical b value using 1.5-T MR. Neuroradiology. 2010;52(8):723-8.

4. Guo L, Liu C, Chen W, Chan Q, Wang G. Dual-source parallel RF transmission for diffusion-weighted imaging of the abdomen using different b values: image quality and apparent diffusion coefficient comparison with conventional single-source transmission. Journal of Magnetic Resonance Imaging. 2013;37(4):875-85.

5. Tachibana Y, Aida N, Niwa T, Nozawa K, Kusagiri K, Mori K, Endo K, Obata T, Inoue T. Analysis of multiple B-value diffusion-weighted imaging in pediatric acute encephalopathy. PLoS One. 2013;8(6):1-10.

6. Gawande RS, Gonzalez G, Messing S, Khurana A, Daldrup-Link HE. Role of diffusion-weighted imaging in differentiating benign and malignant pediatric abdominal tumors. Pediatric Radiology. 2013;43(7):836-45.