Detection and classification of narrow-band gastric endoscopic images by exploring micro-surface structures

• Main Authors: WANG,YI-SIN, 王怡心
• Other Authors: CHANG,HSUAN-TING
• Format: Others
• Language: zh-TW
• Published: 2016
• Online Access: http://ndltd.ncl.edu.tw/handle/93565272001221959647

碩士 === 國立雲林科技大學 === 電機工程系 === 104 === According to the symptoms in microsurface structure of the stomach, the lesion characteristics can be classified into four categories: tortuosity, variation in shape, difference of caliber, and absence of structure. By using the image processing technology on the symptoms appearing in the M-NBI images, we are able to classify the images as either the abnormal or normal ones. In this thesis, we develop an objective computer-aided diagnosis (CAD) system to classify the given M-NBI images. We obtained ethical committee approval and collected a cohort of 100 M-NBI images of gastric cancerous lesions and 10 M-NBI images of normal gastric antrum. We randomly selected 50 abnormal and 5 normal images as the training set to develop our CAD system. The other 55 images were used as the test set. For each image, we selected the region of interest, enhanced brightness, and performed binarization. We then performed a morphological open operation for further extraction of the microsurface structure. To maximize the extraction of microsurface alterations, we carefully utilize 20, 25, 50, and 100 structural element (SE) parameters in the morphological open operation, in which more SE parameters extracted more microsurface features. We then transformed the extracted microsurface features into the values of edge density, in which more complex microsurface features correspond to higher edge density values. Given the training images, we utilize both the artificial neuronal network (ANN) and the support vector machine (SVM) as the classifiers, using the edge density values as the input. The experimental results show the average accuracy rates are 97.76% and 98.05% for ANN and SVM, respectively. The average accuracy rate of SVM classifier is higher than ANN classifier. This method can also be applied to further identify what the symptom is identified when the different symptoms of the microsurface structure are clustered during the training stages in the ANN and SVM. The average accuracy rates of the ANN and SVM are 81% and 80%, respectively.