- 學位論文
一個可偵測多種異常膠囊內視鏡影像之整合辨識系統
An Integrated Recognition System for Detecting Multiple Types of Abnormality in Capsule Endoscope Images
摘要
膠囊內視鏡比起傳統內視鏡更容易觀察到小腸的不正常現象，例如阻塞、出血與腸道白點，但是由於一次檢測就會產生數萬張影像，造成醫師診斷上很大的負擔，因此，本論文旨在研發出膠囊內視鏡影像的自動辨識系統，減少醫師進行診察所需觀看的影像數量，藉此改善醫師在診斷過程花費大量時間的問題，爭取醫療的時效性。 本研究所提出的系統主要是一個四級的辨識器，前一、二級主要針對大面積異常，三、四級則針對形狀不一、分佈零散的小面積異常；第一級辨識器利用HSI轉換篩選黃綠色大面積異常影像(可能是腸道阻塞)；第二級辨識器利用FCM群聚分析，進一步針對第一級所誤判的大面積異常影像(例如出血)進行辨識；之後進入後段辨識器的影像中，多為正常的均勻影像與小面積的異常影像，第三級辨識器的目的是初步篩選出正常的均勻影像；第四級辨識器則使用複雜度較高的倒傳遞類神經網路，主要是偵測出小面積的異常影像(如腸壁白點)。 實驗結果顯示，本系統可以有效篩選膠囊內視鏡影像，減少醫師診斷所需耗費的時間，且平均的異常影像辨識率達到八成以上。此外異常影像多半連續出現，因此即使有幾張零星影像被誤判，也能在其前後的影像中被檢測出。為了盡量不要錯失任一張可疑的影像，因此系統的設計是使異常影像檢測率盡量最高，也因此容易將部分正常影像辨識為異常，但由於系統目的在於輔助診斷，且正常影像數量遠多於異常影像，因此即使有誤判，仍可幫助醫師篩選掉大量影像。
並列摘要
Using a capsule endoscope is better than using the traditional endoscope in detecting abnormal problems of small intestines, such as chime blocked, bleeding, and white dots on the wall of small intestines. However, each examination with the capsule endoscope produces several tens of thousands images, resulting in diagnosis difficulty for a doctor. Therefore, the goal of this thesis is to develop an automatic recognition system for the capsule endoscope images in order to save time for the doctor to view the images and expedite the examination process. The recognition system proposed in this thesis is basically a four-stage classifier; the first and the second stages are aimed at large-area abnormality, and the third and the fourth stages are aimed at irregular–shape, scattered-distribution and small-area abnormality. Here are the descriptions for these four stages: 1.The first stage uses the HSI color model to select the images with large yellow-green abnormal area (could be chime blocked). 2.The second stage uses the FCM clustering analysis to further recognize the images with large abnormal area and are not selected in the first stage (could be bleeding). 3.Most of the images entering the third stage are either normal and uniform or with small abnormal areas. Thus, the purpose of the third stage is to select the normal uniform images. 4.Finally, the last stage uses a relatively more complicated back-propagation neural network (BPNN) to detect the images with small abnormal areas (could be white dots on the wall of small intestines). Experimental results show that the system can do the screening of capsule endoscope images effectively, reducing the time for the doctor to make diagnosis. Moreover, the detection accuracy of the unusual images is up to 80%. Besides, abnormal images usually occurs in sequence. Thus, even a few images are erroneously classified, the abnormality of the small intestines can still be detected by examing their neighboring images. In order not to miss any suspicious image in the detection, the system is designed to make the detection sensitivity for unusual images as high as possible. As a result, some normal images could be determined as abnormal in this way. Since the purpose of this system is to serve as an assisting tool for examination and the number of normal images is usually much more than that of unusual images, the system can still help doctors a lot by screening out many normal images even though the system may make incorrect classification in some cases.