- 學位論文
胰臟腫瘤超音波影像電腦輔助診斷系統之研發
The Development of Computer Aided Diagnosis System for Pancreatic Tumor in Ultrasound Image
摘要
胰臟為人體消化道中重要的器官，卻由於受生理限制讓臨床醫師在進行診斷時具有困難度，而胰臟癌具有極高的致死率。腹部超音波在進行診斷時雖具有很多優點，但在原始超音波影像中容易產生雜訊而導致診斷效果不佳。而本研究之主要目的為研發一套胰臟腫瘤電腦輔助診斷系統，來提供醫師進行臨床診斷之參考意見。 首先將原始超音波影像在消除雜訊、增強對比度與進行邊緣偵測後，可分割出腫瘤之完整輪廓與計算出實際的面積與周長並進行紋理特徵參數分析與形態特徵參數分析。而將分析結果先利用獨立T-test進行篩選後，可找出具有鑑別力的特徵參數並輸入自我組織特徵映射類神經網路中，以不同的設定方式來進行分類。且在對照系統分類之結果與組織病理檢查之結果後，可對系統整體的診斷能力來進行測試與評估。在本研究中共使用40張胰臟超音波影像來進行系統之研發與評估，包括26張腫瘤影像與14張正常組織影像；而在26張腫瘤影像中，則包括9張良性腫瘤影像與17張惡性腫瘤影像。 由實驗結果可發現，系統利用分組訓練與測試的方式來分類腫瘤影像與正常組織影像時，效果會比利用Leave-one-out的方式好，其Accuracy、Sensitivity皆可達到1；且利用6個參考神經元來進行分類時，系統的診斷能力較好，也不會受測試組合影響。系統利用Leave-one-out的方式來分類良性腫瘤影像與惡性腫瘤影像時，其Accuracy可達到0.8462，而Sensitivity可達到1，且利用4個參考神經元來進行分類時，可提升系統診斷能力的Sensitivity。而在超音波影像中，腫瘤形態特徵參數的分類能力會比紋理特徵參數更好。共有8個特徵參數可有效分類腫瘤影像與正常組織影像，而其中4個特徵參數也可有效分類良性腫瘤影像與惡性腫瘤影像。此外，腫瘤的面積則是最重要的形態特徵參數；良性胰臟腫瘤的面積較小，且邊緣也較為平滑；而惡性胰臟腫瘤的面積較大，且邊緣也較為崎嶇。系統處理與分析單張影像所需的平均時間約為25秒。 在本研究中已初步研發出可增強原始超音波影像辨識度與具有特徵參數分析功能之整合式電腦輔助診斷系統，來協助臨床醫師進行診斷並提供參考意見，與降低病患遭誤判或進行侵入性檢查的機率。而系統未來可結合其他醫學影像，來對胰臟腫瘤進行較完整的評估。
並列摘要
The pancreas is one of the important gastrointestinal tract organs. Due to physiological limitations, a physician is hard to make an accurate diagnosis for patients, although the pancreatic cancer had an extreme mortality. The abdominal ultrasound is the most popular way for making a diagnosis; however, a noisy ultrasound image will reduce its overall diagnostic efficiency. The main purpose of this study is to develop a computer-aided diagnosis (CAD) system for pancreatic tumors in ultrasound images to provide a physician some diagnosis information. In this study, after reducing noises, enhancing contrast, and detecting boundary in the original ultrasound image, an entire contour of a tumor was segmented, and its real area and perimeter was calculated to analyze texture and morphological features for this image. After evaluating the results by the independent T-test, the effective features were selected and severd as inputs in the self-organizing map (SOM) which fixed with different modes to classify the ultrasound images. The diagnostic efficiency of this CAD system was evaluated after comparing the classified results of ultrasound images with the pathological results of patients. Totally 40 pancreatic ultrasound images which included 26 pancreatic tumor images (abnormal data) and 14 normal pancreas images (normal data) were respectively used to develop and evaluate this CAD system. Besides, the 26 pancreatic tumor images included 9 benign tumor images (benign data) and 17 malignant tumor images (malignant data). The primary results showed that this CAD system had a better performance by dividing into the training and testing groups than leave-one-out, and its accuracy and sensitivity both were 1 for classifying an ultrasound image as normal or abnormal. The diagnostic efficiency of this CAD system did not affect by grouping when it used 6 reference neurons to classify ultrasound images. When this CAD system classified a pancreatic tumor image as benign or malignant, its accuracy and sensitivity were 0.8462 and 1, respectively. Moreover, Sensitivity of this CAD system was increased when it used 4 reference neurons to classify the pancreatic tumor images. Morphological features had a better performance than texture features for tumor classification in a pancreatic tumor image. In this study, 8 features were proved to classify normal data and abnormal data effectively, and 4 of them were also proved to classify benign data and malignant data effectively. The area of a tumor was the most important morphological feature for tumor classification. A benign pancreatic tumor usually had a smaller area and a smoother contour than a malignant pancreatic tumor. The average time cost for this CAD system is 25 seconds to evaluate an ultrasound image. In this study, the CAD system which combined image enhancement with feature analysis was developed. It could help a physician make a diagnosis, and decreased the probability of making an incorrect or an invasive diagnosis for patients. The CAD system could combine with other medical imaging to make a more complete evaluation tool for pancreatic tumors in the future.