Anomaly Detection For Endoscopic Image

Since gastroscopy is able to observe the interior of gastrointestinal tract directly, it has been widely used for gastrointestinal examination. But it is hard for clinicians to accurately detect gastrointestinal disease due to its great dependence on doctors experi-ences. Therefore, it is very meaningful to analyse the pattern of the endoscopy images and provide computer-aided diagnosis on a large number of images generated in the process of gastroscopy for doctors.The main research contents in this thesis can be summarized as follows:1.The feature extraction algorithms of the endoscopic images. This thesis proposed a feature extraction algorithms which blend the conventional color and texture features and convolutional neural networks features. The experiment results show that our al-gorithms outperform the conventional algorithms at describing the expressions in the endoscopy images which leads to the promotion of the anomaly detection results.2.Cost sensitive anomaly detection technique. This thesis for the characterstics of the data imbalance and different misclassification costs problems, introduced cost sen-sitive factors into the anomaly detection process by using cost sensitive support vector machine to classify the detection units, and proposed sectional area under receiver oper-ating characteristic curve to evaluate the model’s performance and choose the parame-ters in the feature selection algorithms. The experiment results show that cost sensitive detection technique enable the results meet the actual demand of the leaking-detecting probability.