全乳房超音波影像乳腺組織分割與偏移場校正

Abstract

根據衛生署資料統計，乳癌是台灣女性主要死因之一，偵測與預測乳房惡性腫瘤是值得重視的議題。近年來大多使用非侵入式的超音波影像對病人做檢查，尤其在三維全乳房超音波攝影(Whole Breast Ultrasound Imaging)問世之後，它提供了完整的三維乳房解剖構造，有助於醫生為病人診斷。根據研究顯示，乳房密度越高的女性，越容易罹患乳癌。乳房密度計算是預測乳癌的主要指標之一，但由於乳房超音波的雜訊高，三維資料量大，用人工計算乳房密度更顯得吃力，因此本研究將對乳房組織依其音波回音性的高低進行分割。 為了解決超音波影像中模糊的邊界、雜訊高以及灰階質不同質性(intensities in-homogeneity)等問題，本實驗利用改進的等位函數法(Level Set Method)演算法，依照不同組織的音波回音性進行分割。另外，某些區域的乳房組織會被乳頭陰影遮蔽，影響分割的結果。所以本實驗目的除了圈出乳腺組織外，再針對被陰影遮蔽的部分，對Region Based的Variational Level Set Model加入概率的概念做改進，使其能夠反應區域內部灰階值的統計性質，從而取得更準確的分割結果。假設在影像中，局部區域的灰階值呈高斯分佈(Gaussian Distributions)，我們提出改進的基於概率的Level Set Method，此種LSM的能量包含三個變數函數：灰階值平均(mean)，偏移場(bias field)和變異數(variance)。通過能量極小化，使輪廓達到最佳分割。能量極小化是通過交錯更新Level Set Function和求得灰階值、偏移場和變異數的最佳值。模型本身具有計算偏移場(Bias field)的函數，分割的同時可以完成影像校正(Bias Correction)。本論文通過對區域變異數的全域化，進一步改進模型，並通過實驗說明全域變異數和區域變異數的差別。其中全域變異數模型的速度大幅改進。 本實驗採用由台北榮民總醫院提供的U-system全乳房超音波影像，選出正常乳房組織、乳房組織較少和有異常區塊的影像做分割。本實驗的模型，可描繪乳腺組織模糊的邊界，並找出部分被陰影遮蔽較暗的乳腺組織，且對整張影像做偏移場校正。

According to the statistic from Department of Health, breast cancer is one of the leading causes of cancer death among Taiwanese women. Medical research has shown that women with higher breast density have higher incidence of breast cancer. Therefore, estimation of breast density is worth paying more attention. Ultrasound imaging is one of the most important imaging modalities to diagnose breast cancer. Three-dimensional (3D) whole breast ultrasound imaging has been intensively studied, as it provides completed 3D anatomy of breast, which is helpful for diagnosis. However, it is difficult to manually segment the image for the calculation of the breast density, especially for a tremendous amount of data. Therefore, this thesis proposed a model to segment breast parenchyma on whole breast ultrasound images. Blurry object boundaries and inhomogeneity of image intensities on ultrasound images have been a great challenge for automatic image segmentation. In addition, some parts of breast tissues are covered by shadow of nipple, which causes considerable error in the segmentation. To overcome these difficulties, we use a Variational Level Set Method for image segmentation. In our work, the level set method is formulated in a probabilistic model, which is able to take into account the statistical property of the intensities in each region to be segmented. In our improved model, the local image intensities are described by Gaussian distributions, which are used to define the energy functional in a level set formulation. The energy minimization is achieved by level set evolution and estimation of mean, bias field and variances in an iterative process. The bias field and variances of intensities are taken into account in our method for segmentation and bias correction. Furthermore, by globalizing the variance parameter, our model computationally more efficient. This thesis uses U-system whole breast ultrasound images, and these images include the cases with normal parenchyma、less parenchyma and abnormal region. Our model can depict the blurred boundaries of parenchyma 、find parts of parenchyma which is in shadowed area and correct bias field on entire image.