基於定小波轉換及影像特徵分析於電腦斷層影像自動提取耳下腺病灶方法之研究

Abstract

隨著醫學影像技術的進步，要處理大量的臨床資料，電腦輔助診斷已被提出為一有效的解決方法。在近幾年，頭頸部癌症病例數量的增加是種令人憂心的狀況，而耳下腺病灶的出現常伴隨著其他頭頸部區的相關疾病。影像處理被認為是可以實現醫療自動化及輔助診斷的一個有效方法，然而當傳統方法應用於耳下腺部位輔助醫療時，常會遇到因為影像的低對比及其不明顯且模糊的組織邊緣所發生的問題，而使得影像輔助診斷變得困難。所以，在此研究中，為了要針對耳下腺病灶自動化鑑別及定位，提出了特徵式影像分區，並且設計以形態學為基礎的方法以達到提取不同特徵軟組織區域的目的。定小波轉換 (stationary wavelet transform, SWT)，可以在不減低其影像解析之下而提取影像內的局部特徵。耳下腺病灶則是利用所提取出的組織區域其外型幾何特徵作為判號並且再定位。為了更進一步提升病灶區域提取的精確度，主動輪廓模型(active contour model, ACM)則是被用來找尋更精確的組織區輪廓。然而，為了提升主動輪廓迭代在不明顯的組織邊緣上的收斂特性，其初始條件是以繼承組織區域提取的結果，以避免受到外部強邊緣的干擾，而對區域特徵能量圖的強化更可以提升其吸引輪廓演進的效率。根據實驗，其對病灶判斷的真陽性率(true-positive, TP)可達到91%，而偽陽性率為15%。在病灶組織輪廓提取上，其與臨床專家的相似程度可及精確度均可以達到90%以上。

With the progress of medical imaging, computer-aided diagnosis (CAD) is proposed as a solution to deal with the increasing amount of medical data. In recent years, the increasing number of head and neck (H&N) tumor cases is concerned, and the parotid gland lesions may be associated with various H&N disease. To improve medical automation on the parotid, image-based analysis with traditional methods encounters difficulties on the low contrast and weak boundaries of the parotid tissues in the H&N CT images. Therefore, in this research, in order to improve the automation of parotid lesion identification and localization, feature-based segmentation and designed morphological processes are designed to extract the soft tissue regions. Stationary wavelet transform (SWT) is utilized in this work to obtain the local features without resolution degradation. The suspected lesions are identified with their geometric shape features and can be localized. Furthermore, in order to obtain the lesion information with higher accuracy, fine delineation method based on the active contour model (ACM) with automatic initial conditions is proposed. In order to improve the convergence properties on the weak soft tissue boundaries, the initial contour utilizes the segmentation results, and the enhancement of the feature map can increase the attraction on contour evolution. In the experiment, the true-positive (TP) rate of lesion identification can approach 91% and false-positive (FP) rate is about 15%. The lesion delineation can approach average 90% similarity with the results from clinical experts, and accuracy of the fine delineation results on dimension can also approach over 90%.