- 學位論文
電腦輔助診斷系統應用於胰臟超音波腫瘤影像之辨識
The Development of Computer Aided Diagnosis System for Pancreatic Tumor Recognition in Ultrasound Images
摘要
胰臟因生理位置的因素，導致胰臟炎、胰臟囊腫及胰臟癌皆不易診斷。不同的胰臟腫瘤其致死率、嚴重性、治療方法和其效果也皆不盡相同。而腹部超音波是目前常見的診斷方法，雖然此方法有許多優點，但在成像階段易產生雜訊而影響醫師的診斷。因此，本研究的主要目的就是：研發超音波影像之電腦輔助診斷系統用於胰臟腫瘤的辨識，輔助醫師做出正確的診斷。 首先對原始超音波影像進行增強對比度的前處理動作，接著分割出腫瘤的完整輪廓及影像，並擷取腫瘤的紋理和形態特徵進行T-test分析得到具鑑別力之特徵參數，在將這些參數輸入類神經網路對腫瘤影像做分類，最後對照病理檢查的結果來對系統做評估。在本研究中共使用69張胰臟超音波影像來進行系統之研發與評估，其中包括13張正常組織影像與56張腫瘤影像；而其中腫瘤影像又包括了18張實質惡性腫瘤影像、6張良性偽腫瘤影像以及23張易癌變囊腫影像、9張發炎性偽囊腫影像。 由實驗結果可發現，面積(Area)及正規化徑向長度之平均面積(NRL_MA)是所有特徵參數中最具鑑別力的兩個，和醫學上腫瘤面積大小常用來判斷腫瘤良惡性之觀念相符。此外，將輸入類神經網路進行分類之影像的訓練組與測試組比例設定為50%-50%時的系統效能會優於比例是80%-20%時，且在進行不同階段的分類時使用不同數量的參考神經元來進行也可提升系統的診斷效能：在第一階段分類正常組織及腫瘤影像時，以6個參考神經元的效能最佳其中Accuracy可達0.957；而第二階段分類腫瘤與囊腫影像時則是4個參考神經元的效能最佳，其中Accuracy也有0.946；最後第三階段的腫瘤分類及囊腫分類時，又都是以6個參考神經元的效能為最佳，Accuracy分別達到0.917、1；另外，在某些分類階段使用原始超音波影像來做為判斷標準亦可提升系統之效能。 本研究中已研發出可增強原始超音波影像辨識度及具有特徵參數分析功能之電腦輔助診斷系統，來協助臨床醫師進行診斷並提供參考，以降低病患遭誤判或進行侵入性檢查之機率，而未來系統亦可結合其他不同之胰臟腫瘤影像，來對胰臟腫瘤進行更完整的分析及研究。
並列摘要
Due to the anatomic position, it is hard for physicians to diagnose pancreatic cancer. The different pancreatic tumors whose fatalities, treatments, recoveries are all different. Nowadays, the abdominal ultrasound is a common way to diagnose pancreatic tumors, which has a lot of advantage; however, it is apt to get a lot of noises when imaging; as a result, it will impact the diagnoses. Therefore, the purpose of this study is to develop a computer-aided diagnosis (CAD) system for pancreatic tumors in ultrasound images in order to help the doctors to make the diagnoses more precisely. First of all, we enhanced the contrast of the US images as a preprocessing; afterwards, we segmented the area of the tumors for analyzing the textural and morphological features by T-test. Subsequently, enter the significant features into the self-organizing map (S.O.M) for categorization. Lastly, compare the categorized outcome with pathological result. There were 69 US images which were used in this study, including 13 images of normal tissues, 18 images of adenocarcinoma tumors, 6 images of pancreatitis pseudo tumors, 23 images of cystic tumors, and 9 images of pancreatitis pseudo cysts. The result showed that Area and NRL_MA are the most two significant features among all, which correspond with the theory that the larger tumors are likely to be malignant. Besides, the system performance was better by entering both 50%images into SOM for training and testing than by entering 80% images for training and 20% images for testing. Furthermore, it is feasible to improve the system performance by applying different number of competitive neurons in different stages. In the 1st stage, for identifying tumors from normal tissues, the accuracy can reach 0.957 by applying 6 competitive neurons; in the 2nd stage, for identifying cysts from tumors, the accuracy can reach 0.946 by applying 4 competitive neurons; in the 3rd stage, for identifying tumors and cysts, the accuracy can reach 0.917, 1, respectively, by both applying 6 competitive neurons. In addition, using the original US images instead of the preprocessed ones can improve the system performance in some stages as well. In the study, the CAD system combined image enhancement with features analyzing function was developed. It can provide physicians with more specific information, reducing the misdiagnosed cases. In the future, we hope this CAD system can apply with other kinds of pancreatic tumors images for the sake of thoroughly researching pancreatic tumors.