結合動態對比顯影與擴散磁振影像特徵分析於自動化乳房腫瘤劃分之研究

Abstract

近三十年，癌症連續蟬聯國人十大死因榜首，其中發現乳癌於女性罹癌率從民國八十四年的第五位，至九十五年已爬升至首位，且其罹癌率仍在持續攀升中，因此如何能夠發展一有效之乳癌篩檢方法是急切需要的。自1980年代發展至今，磁振造影(Magnetic Resonance Imaging, MRI)已成為相當重要的醫學影像檢查技術之一，可提供施測者的相關病理資訊，供臨床醫師進行術前評估、治療療效評估與術後追蹤。進行乳房MRI檢查時，一般會配合注射顯影劑於靜脈，針對不同時間點的顯影訊號，利用快速造影技術來取得連續之影像並觀察腫瘤區域之訊號變化，藉以評估腫瘤之特性，此方法稱為動態對比顯影(Dynamic Contrast Enhanced-MRI, DCE-MRI)。目前MRI常用於乳房腫瘤診斷上的主要兩項影像技術，分別為DCE-MRI及擴散係數(Apparent Diffusion Coefficient, ADC)量測，上述技術雖然目前已是臨床上之標準序列，但由於水分子之擴散在腫瘤區域內容易受到組織型態之改變造成擴散現象並非以高斯性呈現，擴散係數之準確性亦受其影響，有鑑於此，應用擴散峰度成像(Diffusion Kurtosis Imaging, DKI)來觀察乳房腫瘤之擴散非高斯性已成為臨床上急欲發展技術之一，初步結果也顯示DKI能夠提供較擴散係數為佳之敏感度。臨床醫師對於腫瘤區域劃分之判斷主要根據DCE-MRI與ADC之資訊，但本研究更進一步地將DKI之非高斯性資訊也列納入腫瘤自動劃分之演算法中，期能透過更詳盡的特徵來提升腫瘤自動化劃分程序的穩定性(Robustness)及準確性(Accuracy)。本研究所開發之演算法首先以k-means方法對DCE-MRI影像作訊號分群，接著計算由臨床醫師手動圈選腫瘤區域內的ADC及DKI影像參數資訊並進行直方圖分析以設定腫瘤篩選條件，並計算腫瘤可能性，且選取腫瘤可能性較高的區域作為腫瘤化分區。本研究開發上述演算法，並應用在臨床所收集的十四組資料中，結果顯示本自動化腫瘤劃分法與臨床醫師圈選之腫瘤區域相似度可達60.16%，未來，此一自動化腫瘤劃分法將有機會提供臨床上進行腫瘤區域判讀之協助。

Since the past 30 years, cancer has been ranked first among the top ten leading causes of death. Among all the cancers, breast cancer was ranked fifth for women in 1995 and became first in 2006. To date, the incidence rate of breast cancer is still increasing. Therefore, a cost-effective and early diagnostic medical imaging tool for breast cancer is highly needed. Since 1980, magnetic resonance imaging (MRI) has been considered as an important medical imaging modality for clinical applications. MRI could provide useful pathological information for assessing the preoperative evaluation, efficacy of the treatment and postoperative tracking. On clinical protocols, breast MRI is usually performed after administration of contrast agents into the blood stream for assessing the perfusion information in microenvironment. To investigate the dynamic changes of signal intensities for a given region-of-interest (ROI), fast T1-weighted imaging technique is typically utilized to retrieve continuous signal changes, which is known as dynamic contrast enhanced-MRI (DCE-MRI). Another MRI technique used for breast tumor diagnosis is diffusion MRI, which can produces a quantitative metric, namely, apparent diffusion coefficient (ADC). In tumor regions, complex tissue microstructure and microenvironment typically cause increased diffusion heterogeneity, that is, non-Gaussianity. Therefore, a recently developed diffusion MRI technique, diffusion kurtosis imaging (DKI), is used to assess diffusion non-Gaussianity of water molecules in breast tumor regions. In this study, assessment of diffusion non-Gaussianity by DKI will be taken into account for tumor segmentation, except for DCE-MRI and ADC measurement which are usually applied to differentiate the tumor lesion by radiologists. By incorporating more information into our segmentation framework, we aimed to increase robustness and accuracy for automatic breast tumor segmentation. Our algorithm first performs k-means clustering approach on DCE-MRI images to determine the initial regions for following segmentation. Secondly, the histogram analysis is performed on ADC and DKI images to determine tumor probabilities for identifying possible tumor ROI. In results, a total of 14 clinical cases have been examined by the developed approach, showing the similarity between automatically selected ROI and manually selected ROI is approximately 60%. In future, this proposed segmentation approach could be potentially helpful to facilitate the early diagnosis and evaluation of treatment effects on breast cancer.