Neural Network Learning Method For X-ray Medical Image Detection In Small Sample Environment

With the rapid development of artificial intelligence,the research on medical image classification and image detection methods based on deep learning neural networks is also in progress.The use of deep learning neural network methods for target detection and annotation of X-ray films acquired by traditional methods is also a hot research topic.Currently,the study of neural network learning methods for X-ray medical image detection in a small sample environment is being carried out to address the problems of difficult X-ray image detection and inaccurate localization.The main research elements are as follows:(1)Just-in-time learning method for image anomaly detection based on time slot transferring.The training and updating of the model in the new device is aimed at the lack of labeled samples.Firstly,the samples of the old device are divided into time periods according to the working conditions of the device;secondly,a small number of samples of the new device in each time period is used as a template to select a similar sample set in the corresponding time period of the old device;then,the obtained sample set and the small number of samples of the new device are jointly used to update the model parameters of the new device;finally,the method in this chapter is verified in the simulation of the example environment to be superior to the direct migration learning of the model of the old device parameters.(2)Faster R-CNN image target localization method based on feature enhancement.For target localization of anomaly image detection in work(1),firstly,a structurally optimized Faster R-CNN model is built: a Multi-scale Retinex preprocessing module with enhanced sample texture,a feature extraction module connected by residual blocks,and a fully connected layer based on Dropout regularization;secondly,the optimized model is used to train in the new device;then,the model is used for image target online Then,the model is used for online localization of image targets,and the model is evaluated with accuracy as a metric;finally,the method is verified to be better than the original Faster R-CNN algorithm in an example environment simulation.(3)Non-linear real-time filtering-based detection model parameter update method.In view of the uncertainty of equipment working condition changes,which makes the obtained image samples contain large noise,firstly,the model pooling layer is updated using ROI Align to improve the anti-interference capability of the model;secondly,the optimized model parameters are established as a dynamic model;then,a non-linear Kalman filter-based network parameter updating method is established;finally,the parameters of the detection model are used to update,and in the example environment simulation to verify that this method is superior to the gradient descent method.In summary,this paper,in work(1),the problem of inapplicability due to direct migration learning of model parameters from old equipment to new equipment models is solved;in work(2),the problem that image samples with feature differences are difficult to be detected and located by the mode is solved;in work(3),the problem that the originally used gradient descent method cannot meet the adaptive updating of model parameters in an environment with changing working conditions is solved.