運用決策樹於腦部核磁共振影像之監督式腦組織分割研究

Abstract

本文研究之目的主要是運用決策樹以增進腦部核磁共振影像於組織分割的準確度達到一個高品質的腦部組織分割為目標，首先，此研究提出用一個多階層仿視覺 (Multiscale retinex) 演算法來校準核磁共振影像之不均勻度 (Inhomogeneity)，此實驗之核磁共振影像含有假體核磁共振影像與老鼠之腦部核磁共振影像，經此演算法校準後可獲得較好的影像品質與清清晰的深層腦部結構影像，此外也用訊號峰值對雜訊比 (Peak-signal-to-noise) 與 (Contrast-to-noise) 對比度對雜訊比來評估校準後的影像品質。進一步實驗為得到腦部不同組織解剖結構，我們提出用分類與迴歸決策樹 (CART) 來分割模擬之假體核磁共振影像與模擬之腦部核磁共振影像，進行組織分割實驗後得到良好的腦部解剖組織影像。接下來更進一步實驗為增加腦部組織分割的準確度以提供神經解剖應用性，我們提出用一個促進式之決策樹 (Boosted decision tree) 分割模擬之假體核磁共振影像與模擬之腦部核磁共振影像與人體實體之腦部核磁共振影像，進行組織分割實驗後得到更準確之腦部解剖組織影像。以上兩個實驗之分割結果並且都用準確率 (Accuracy rate) 與k指標 (k index) 來量化評估其影像分割之表現。最後一個實驗是要得到腦部核磁共振影像之高品質分割影像，我們提出用一個促進式之決策樹為基礎結合以一個多階層仿視覺演算法為影像前處理之不均勻度校準分析，接著用促進式之決策樹來分割模擬之腦部核磁共振影像進行組織分割，除了能成功地對模擬之腦部核磁共振影像進行組織分割之外，經實驗分析後得到非常準確之腦部解剖組織影像，大大地改善了核磁共振影像之不均勻度與雜訊對影像分割之準確度所造成之影響，更加提高了組織分割的準確率，最後結果得到高品質的腦部組織分割，結果將對於腦部神經解剖之應用性有很大的幫助。

The purpose of this research is to achieve the high quality of brain tissue segmentation by using decision trees from magnetic resonance (MR) images. First, we proposed a multiscale retinex algorithm to correct the intensity inhomogeneity of brain MR images which include phantom data and animal data to obtain better image quality and clearer deep brain structures. The intensity inhomogeneity often occurs in MR images which are received by surface coils. The peak-signal-to noise (PSNR) and contrast-to- noise (CNR) were both used to evaluate the correction performance. Second, we used a classification and regression tree (CART) to segment brain MR images, including simulated phantom MR (SPMR) images and simulated brain MR (SBMR) images, to successfully segment brain tissues. Third, we proposed a boosted decision tree combined with fuzzy threshold to segment brain MR images, including SPMR images, SBMR images, and a real data, to achieve higher accuracy rate of brain tissue segmentation. The accuracy rate and k index were better when we used the boosted decision tree algorithm combined with a fuzzy threshold to classify gray matter (GM), white matter (WM), or cerebral-spinal fluid (CSF) in brain. Finally, we used the boosted decision tree through preprocessing procedure by the multiscle retinex algorithm to greatly improve the brain tissue segmentation from SBMR images. The accuracy rate was the best when we used the boosted decision tree algorithm combined with a multiscale retinex algorithm as a preprocessing procedure to classify GM, WM, or CSF in brain MR images. The results shows that a boosted decision tree combined with preprocessing by the multiscale retinex algorithm can successfully improve the accuracy rate of MR brain tissue segmentation and can achieve a high quality of tissue segmentation in brain MR images.