三維腫瘤分割技術應用於正子攝影影像

Abstract

正子攝影(Positron Emission Tomography, PET)影像由於是非侵入式的、高篩檢率，所以在現今醫療的診斷上佔有重要的地位，且成為常見的診斷工具之一。然而在製作放射治療計畫時，醫生必須付出大量人力圈選腫瘤，為了減省不必要的人力花費，自動化分割腫瘤是必要的，然而在正子攝影影像分割上由於低解析度、高度變化的像素值以及低信雜比，因此造成分割上的困難。也因為難以物體的邊緣及背景具有相似的像素值，因此難以區分兩者，導致腫瘤分割上的困難，也會對醫生的診斷造成影響。其中，小腫瘤影像由於資訊量少，因此腫瘤邊緣經常被分類為背景，造成分割不足的情形，另外低對比度影像由於腫瘤與週邊組織太過相近，因此容易造成過度分割的情況。本論文針對肺部腫瘤提出一適應性調整VOI大小及對於小腫瘤及低對比度影像影響程度低的統計型分割演算法，利用峰度值直方圖中最似高斯的兩個區段內資訊匹配高斯曲線，利用所得兩曲線得到最佳閥值，此方法可由調整VOI大小減少背景雜訊的干擾，且能利用峰度找到最佳高斯分布的範圍，進而得到較佳的分割效果。本演算法的分割結果與醫生所圈選的參考值平均有80.38%的相似度，在運算時間上也比大部分演算法快速。

With the non-invasive and high screening rates, Positron Emission Tomography image plays an important role in today's medical diagnosis and becomes a common diagnostic tool. However, in the production of radiation treatment plan, the doctor must spend a lot of effort determining the region of tumor. In order to reduce unnecessary labor costs, automated segmentation is significant to clinical diagnosis. It is difficult to segment PET images due to the low resolution, high variation of intensity and low signal to noise ratio. It is hard to detect the boundary between background and tumor with similar intensity as well. Especially for the image with small tumor, there is only a little hot spot in image. The fact that the edge of small tumor is often recognized as background causes the situation of insufficient segmentation. On the other hand, for the low contrast images, the background is often classified as a tumor and it causes the situation of over segmentation. It is because the intensity of tumor and surrounding tissue are too similar. We propose an algorithm for lung cancer which can adaptively adjust the size of VOI and low-interference with small tumor image and low contrast image. We use Kurtosis value to find the best two ranges of histogram which have the biggest similarity to Gaussian distribution. This algorithm can reduce the interference of background noise by adjusting the size of the VOI. It can also take advantage of Kurtosis to get the best distribution and then find the best threshold. The similarity between given result and ground truth defined by doctor reaches 80.38%. It also cost less time than most algorithms.