膠囊內視鏡對於消化道檢查帶來許多利多，然而後續的影像判讀卻是個大 工程，因為病人的影像將高達5 萬多張，一位熟練的專業醫師大約需耗費2~3 小時來查看一位病人的影像，造成很大的負擔，因此建構一套膠囊內視鏡影像 的分群系統，有其必要性。過去文獻中大多使用監督式方法分類病炤成二群， 此方法分類前需要提供訓練資料，再利用這些資料套入訓練模型，以訓練出最 佳分類規則，其缺點為受限於訓練資料的樣本取得，且每個樣本需要有準確的 類別標籤，再加上膠囊內視鏡影像可能內含多種類別，而且需要訓練時間，因 此膠囊內視鏡影像的分類較不適合使用監督式方法，本研究將使用非監督式的 方法來分群膠囊內視鏡影像，先計算膠囊內視鏡影像之灰階強度統計圖、灰階 共生矩陣、與區域二元樣式 (local binary pattern, LBP) 等特徵，經過特徵選擇 後，再使用K 平均法 (K-means)、模糊C 平均法 (Fuzzy C-means)、與自我組 織映射 (self-organizing map, SOM) 來群聚正常、出血、潰瘍等小腸的樣式， 最後針對分群績效進行評估，計算分群之精確性 (Precision BCubed) 與召回性 (Recall BCubed) 二個指標。實驗影像數據由PillCam Capsule Endoscopy Images Atlas 取得，實驗結果發現，當使用資訊獲利取其前10 個特徵後，再經過標準 正規化程序，接著利用兩階段之分群方式，第一階段採用K-means 分離出出血 與其他影像，第二階段再採用K-means 或Fuzzy C-means 分離出正常與潰瘍影 像，如此可獲得足以媲美監督式分類法之分群績效。

Wireless Capsule Endoscopy (WCE) is a new technology from a traditional endoscope to detect abnormal patterns in the small intestines,such as bleeding and white dots (ulcer). Each examination with the capsule endoscope produces about fifty thousand images, resulting in diagnosis difficulty. Typically, a medical clinician spends more than two hours to analyze a WCE video. Research has been attempted to automatically find abnormal regions to reduce the time needed to analyze the videos. Most of these methods have been based on supervised classification to distinguish two classes. This method requires training data to feed into a learning model for obtaining the best classification rules. Thus, the difficulties of supervised methods include: limiting to sample the training data, accurate category labels, and extended training. Therefore, classification of capsule endoscopy images is less suitable for use supervised methods. The present work is to develop a recognition system for the capsule endoscope images based on unsupervised methods such as K-means, Fuzzy C-means, and Self-Organizing Map (SOM). Features are obtained from Gray-Level Histogram (GLH), Gray-Level Co-occurrence Matrix (GLCM), and Local Binary Patterns (LBP). We compute the Precision BCubed and Recall BCubed to compare the performance of the cluster methods. WCE sequences will be downloaded from PillCam capsule endoscopy images Atlas. Experimental results show that the optimal clustering method for WCE sequence analysis is as follows. First, the information gain is used to obtain the best 10 features. After the standard normalization procedure, the features are then clustered by two phases. The first phase applies K-means to separate bleeding and other images. The second stage can then use K-means or Fuzzy C-means to distinguish normal and ulcer images. The results are comparable to the performance of supervised classification methods.