Research On Deep Learning-based Anomaly Detection Methods Of Esophageal Endoscopic Images

Due to differences in dietary habits and lifestyles,the incidence of esophageal diseases,such as esophageal cancer,in China is far higher than the world average.Endoscopy of esophageal is the most effective way to examine and diagnose esophageal diseases.However,the number of professional endoscopic physicians is limited,and relying solely on doctors for large screening of esophageal endoscopic images is time-consuming and expensive.Moreover,traditional deep learning based computer aided diagnosis techniques require large-scale datasets with sufficient quantities and balanced categories for training,which is difficult to achieve in the field of medical imaging,while anomaly detection can solve this problem.Anomaly detection is an important field in machine learning,which refers to detecting abnormal data from a large amount of normal data.The training set usually only contains normal samples,and thus it is actually a binary classification task under semi-supervised conditions.In this thesis,we focus on anomaly detection algorithms for esophageal endoscope images,treating healthy esophageal images as normal samples,and constructing an anomaly detection model to detect diseased esophageal images under the condition that the training set contains only normal samples.The main contributions of this thesis is as follows:(1)This thesis provides a detailed review of the development of anomaly detection.Based on whether deep learning techniques are used,anomaly detection methods are divided into traditional anomaly detection methods and deep learning based anomaly detection methods.Representative algorithms in both categories are systematically summarized and reviewed,and the research progress of anomaly detection in the field of medical imaging is introduced.(2)Following the basic idea of reconstruction-based anomaly detection methods,we use a Variational Auto-encoder as the framework,and adopt skip connection technology to achieve the transmission of low-level semantic information.We also introduces a memory module to suppress the generalization ability of the model by recording the feature vectors of normal samples,enhancing the difference in reconstruction error of normal and abnormal samples,thereby achieving better anomaly detection results.(3)To further improve the performance of memory module,we add an attention mechanism to the memory module,and introduce a clustering algorithm to improve the distribution of memory vectors in the feature space.We treat memory vectors as clustering centers and cluster the original feature vectors to be recorded.A new clustering loss function is proposed by evaluating the clustering effect using the scatter matrix.The clustering algorithm proposed in this thesis can make the memory vectors record original feature vectors more accurately,and be widely distributed in space to record more features,thereby improving the performance of the memory module.(4)Currently,the available public datasets of esophageal endoscope images are limited.Therefore,we carefully select and preprocess esophageal endoscope images collected from two hospitals to construct two esophageal endoscope image datasets.Each image in the dataset is labeled by two professional endoscopic physicians to ensure its accuracy.Our proposed model conduct extensive comparative and ablation experiments on the two datasets,demonstrating its excellent performance in anomaly detection of esophageal endoscope images.