- 學位論文
胰臟腫瘤斷層掃描影像電腦輔助診斷系統之研發
The Development of Computer Aided Diagnosis System for Pancreatic Tumor in CT Image
摘要
摘要 胰臟為人體消化道中之重要器官，致死率極高的胰臟癌不易早期偵測，因此臨床上常以有施打顯影劑之影像進行診斷。本研究利用影像處理技術，開發一套電腦輔助診斷系統以協助醫師進行早期偵測，並且探討降低病人施打顯影劑的可能性。 首先以中值濾波器與小波轉換法對原始CT影像做前處理；再分別利用直方圖等化法及Otsu’s二值化法得到胰臟的影像範圍；再以區域成長法將分割出腫瘤與胰臟影像；最後將ROI分成「影像處理後胰臟範圍」、「影像處理後腫瘤範圍」、「原始影像胰臟範圍」、「原始影像腫瘤範圍」四種不同的子影像並進行紋理特徵參數分析與形態特徵參數分析，以探討(1)影像處理過程是否破壞影像中重要的生理資訊(2)何者對胰臟的影像範圍與腫瘤的影像範圍會有較準確的影像分辨能力。研究中使用有施打顯影劑之CT影像41張；未施打之CT影像36張。利用獨立T-test篩選出有鑑別力的特徵參數，並且輸入自我組織特徵映射類神經網路(SOM)中進行分類，對照系統分類結果與組織病理檢查結果，評估系統整體的診斷能力。 結果顯示: 1. 有施打顯影劑之影像部分，腫瘤組織與胰臟組織的分類(第一階段分類)使用「影像處理後胰臟範圍」有最好的分類效能，其Sensitivity=0.973；Accuracy=0.873；Kappa=0.839；腫瘤良惡性的分類(第二階段分類)使用「原始影像腫瘤範圍」有最好的分類效能，其Sensitivity=0.897；Accuracy=0.914；Kappa=0.748，系統單次影像分類所需的平均時間約13秒。2. 未施打的部分，第一階段分類使用「原始影像腫瘤範圍」有最好的分類效能，其Sensitivity=0.994；Accuracy=0.995；Kappa=0.984；第二階段分類使用子影像「原始影像腫瘤範圍」有最好的分類效能(Sensitivity=0.874；Accuracy=0.887；Kappa=0.583)，系統單次影像分類所需的平均時間約11秒。此外在選取良惡性判讀特徵上 對有施打顯影劑之影像其惡性腫瘤有較低的Energy，而在無顯影劑之影像，以Entropy、Energy這兩個紋理特徵參數最為明顯，這些可提供醫師進行診斷之參考。 本研究已初步研發出可增強CT影像的辨識度，並且具有特徵參數分析之電腦輔助診斷系統，能協助醫師進行診斷並提供參考意見。目前系統對有無顯影劑之影像的診斷效能相近，尚需要更多的影像資料進行評估，以確定病患施打顯影劑的必要性。
並列摘要
Abstract The pancreas is the important organ in the digestive tract of human body. The pancreatic cancer has an extreme mortality, because it is hard to be detected in the early stage for physicians. Therefore, enhanced images that injected contrast medium are often used in the clinical diagnosis. This study will develop a computer-aided diagnosis system by applying image processing to help physicians detect pancreatic cancer in the early stage, and explore the possibility of reducing the patient injected contrast medium. First of all, the original CT image was preprocessed by using the median filter and wavelet transform. Secondly, histogram equalization and Otsu’s method were used to get the image region of pancreas. And then, the region growing method was used to divide an image into the images of pancreatic tumor and pancreas. Finally, ROI of images was divided into four different sub-images as “the region of pancreas in processed image”, “the region of pancreatic tumor in processed image”, “the region of pancreas in original image” and “the region of pancreatic tumor in original image”. The sub-images are proceeded with textural features analysis and morphological features analysis to explore whether the important physiological information of the image are lost during image processing, and which is the more accurate system to distinguish between the region of pancreas and pancreatic tumor. Totally, 41 enhanced CT images and 36 unenhanced CT images are used as train and test data sets. The features have been chosen with independent T-test and entered into self-organizing map (SOM) to be classified, and then the result of system classified contrast with the pathological result of patients to evaluate the developed system. The results show as following. 1. For enhanced CT images, using “the region of pancreas in processed image” is the best way to classify the tissue of tumor and pancreas (the first stage classification) with sensitivity=0.973, accuracy=0.873, kappa=0.839, and using “the region of pancreatic tumor in original image” is the best way to classify tumor as benign or malignant (the second stage classification) with sensitivity=0.897, accuracy=0.914, kappa=0.748, respectively. The average time cost for the classification of single image is about 13 seconds. 2. For unenhanced CT images, using “the region of pancreatic tumor in original image” is the best way both in the classification of first and second stage with sensitivity=0.994, accuracy=0.995, kappa=0.984, and sensitivity=0.874, accuracy=0.887, kappa=0.583, respectively. The average time cost for the classification of single image is about 11 seconds. Moreover, the malignant tumor has lower value of “Energy” in the enhanced CT images and the two textural features as “Entropy” and “Energy” are the most notable in the unenhanced CT images on selecting the characteristic interpretations of benign or malignant tumors, which provide physicians as diagnosis reference. This study has already developed a computer-aided diagnosis system for detecting pancreatic cancer in the early stage, and it could help physicians and provide them second opinion. So far, there are no significant differences for the performance of system for unenhanced and enhanced CT images, however, we still need more data to evaluate and confirm the necessity of injecting medical contrast medium into patients.