Study Some Key Issues On Gastric Image Understanding Based On Convolutional Neural Networks

Gastroscopy is the prior method for upper gastrointestinal lesions screening.It uses the tubular gastroscopy equipment to sequentially check the actual conditions of the mucosa of the esophagus,stomach,and duodenum.Gastroscopy can screen and take high-quality gastroscopic images of the lesions.The doctor interprets these images based on experience to achieve an intuitive clinical diagnosis.In order to help doctors to find lesions and improve the diagnosis efficiency,many researchers use the computer-aided detection system,which takes image understanding as the core technology.Traditional medical image understanding algorithm realizes the recognition and detection of medical images by extracting the concerned features,and based on the features,build the feature classifiers.However,researchers need to manually design features based on the prior knowledge of specific fields and tasks,which takes much time and effort.And because of the subjective influence of feature selection,the stability and the generalization of the model are difficult to guarantee.In recent years,with the development of computing capacity,optimization algorithms and high-quality medical images,deep learning,especially Convolutional Neural Network(CNN)has been widely used in medical image analysis.CNN is an end-to-end architecture which can extract and interpret deep features from datasets.Deep learning has gradually replaced the traditional machine learning algorithms and become the mainstream method of medical image understanding,and gastroscopy is no exception.However,the following issues still need to be resolved in understanding gastric images based on deep learning: 1)Although the number of gastric images is huge,the number of disease images is small,which may cause model over-fitting.2)Most lesions appear as small targets,and the existing object detection networks are not ideal for small object detection.3)There is a strong spatial relationship between diseases and anatomies in the field of gastroscope image,but the conventional networks do not make full use of this prior knowledge.4)Datasets for different tasks are usually labeled independently.The model trained on one dataset is usually only for a single target task.However,in order to meet the actual clinical application,it is usually necessary to build a multi-task-oriented fusion model.5)Because the scene of gastroscopy is relatively complex,the data of the training model cannot reflect it well.In order to solve the problems above,this thesis conducted theory and method research on white-light gastric image understanding based on CNN,and we have built a real-time gastric video screening system.Specifically,this thesis mainly includes the following aspects:1)A method of training from scratch based on iteration for concise model was studied.Due to the large amount of parameters in the conventional network,it is easy to overfit.Therefore,we introduced the fire module from Squeeze Net to reduce the model parameters and improve the time performance.At the same time,in order to avoid the introduction of external datasets,this thesis proposed an iterative non migration training method,which improved the accuracy by retraining small parameters several times.The experimental results showed that the proposed method could achieve considerable accuracy compared with the transfer learning method with the help of pretraining model.Moreover,compared with the transfer learning method,the accuracy of this method is improved by 9.16% to 88.9% in the application of small dataset for gastric precancerous diseases.2)The real-time object detection algorithm based on multi-scale features and pooling layer fusion was studied.The conventional object detection networks can reduce the parameters and improve the generalization ability of the model by maxpooling layer,but it will also lose detail information,which is not good for small object detection.This study improved the conventional object detection networks:(a)the deconvolution module was introduced to improve the ability of image feature extraction.(b)In order to reuse the lost information caused by pooling layers,we proposed the fusion of the multiple pooling methods to replace max-pooling layer.The algorithm proposed in this thesis could improve the recall of the gastric polyp detection(especially small lesions)and reduce the rate of missed detection.Our model improved mean average precision(m AP)by nearly 2%.Although our network sacrificed the time performance,it could still detect upto 50 FPS.3)The multi-task learning of object detection algorithm based on graph convolutional networks(GCN)was studied.Conventional object detection algorithms usually focus on extracting the features of the objects and ignore the relationship between the objects(for example,in the field of gastroscopic images,antrum and pylorus usually appear in pairs).In order to fully use this domain knowledge,GCN module was introduced.GCN could express the dependence of each anatomical structure and participate in the detection process of anatomical structure as an auxiliary task.The experimental results show that the method can reduce the false-positive samples and improve the m AP of the model by 0.87%.4)The model fusion algorithm based on the selection of conditional negative samples was studied.The annotation of datasets for different target tasks is usually carried out independently.Different tasks will train different models.In clinical scenarios,multi-task oriented fusion model is often needed to improve the time performance.In this thesis,the fusion model algorithm was based on default boxes in detection networks.It combined offline with online training methods and as far as possible to eliminate the false negative samples in the default boxes,which could reduce the probability that false negative samples became difficult-training negative samples.The algorithm trained a fusion model which could detect disease and anatomies simultaneously without relabeling the dataset and without losing the detection accuracy.Compared with non-fusion model,the time performance of the fusion model could be improved by 33.53 FPS.5)In this thesis,a real-time gastric screening system based on CNN was built.Time performance of the system could reach 43.13 FPS.Through retrospective analysis of clinical data,it was found that the detection rate of the system have reached the level of clinicians.