Research On Key Techniques Of Medical Image Diagnosis Based On Deep Convolution Neural Network

At present,convolutional neural network models have been widely used to solve various traditional problems,such as face recognition,text recognition,identification/license plate recognition.Its powerful ability to solve computer vision problems has been verified in multiple natural scene tasks.Medical image-assisted diagnostic scenarios are complex scenes involving multiple types of images.These include:ordinary two-dimensional images,super-resolution images,and three-dimensional images.Whether it is possible to migrate existing successful models,process different types of medical images and draw correct diagnosis conclusions,and whether the model processing ability can meet clinical use requirements is a problem.This paper summarizes the research progress of convolutional neural networks in the field of medical imaging diagnostic model design,aiming at adapting different data formats,migrating existing models,optimizing specific application scenarios,balancing the speed,accuracy and interpretability of models.Research on issues such as improving the usability of the model.A general two-dimensional medical image data processing method was proposed,and a Trans-DRNet model was constructed for the diabetic retinopathy detection scene.The model draws on the doctor's actual diagnosis and treatment process,and completes the effective fusion of the characteristics of both eyes through a two-stage convolutional neural network,which better characterizes the severity of the disease.In addition,the advantages and disadvantages of different solving methods are analyzed in the model solving process,and the applicable scenarios of different solving methods are expounded.Experiments show that the model has strong classification ability of lesion images.A super-resolution medical image data processing method was proposed,and a Trans-TumorNet model was constructed for the sentinel lymph node metastasis detection scene of breast cancer.The model combines the convolutional neural network on the microscopic scale with the morphological calculation on the macroscopic scale,and constructs the complete chain of tumor region information extraction,which has strong tumor image classification ability.In addition,aiming at how to balance the model efficiency,model performance and model interpretability,this paper proposes a method named MMFE that can utilize multi-resolution features for the improvement of Trans-TumorNet model by carefully observing the process of doctor reading.A three-dimensional medical image data processing method was proposed,and the Trans-LungNet model was constructed for the early lung cancer screening scene.The model uses two convolutional neural networks with completely different structures to extract nodule features and special tissue features,and complete the structural processing of 3D images.Experiments show that the model has strong 3D image classification ability.In addition,in terms of time dimension expansion,this paper explores the application of object detection model in breast ultrasound nodule detection scenario by proposing Trans-USNet model.By decomposing the problem,the model divides the nodule detection and nodule classification into two stages,which take into account the speed requirements of video processing and the accuracy requirements of clinical diagnosis.Experiments show that the model can process video data efficiently and has strong nodule detection and classification capabilities.A general method for improving the accessibility of convolutional neural network models is proposed.This paper points out that usability improvements need to improve the interpretability and usability of the model.For the interpretability improvement of the model,a method based on classification gradient mapping is proposed to mark the source regions of the features that significantly affect the diagnosis results.For the improvement of the usability of the model,a method of encapsulating the model as a service by using WebService and asynchronous message queue is proposed.At the same time,in order to verify the validity of the method,this paper improves the Trans-DRNet model by using this method,and creates a reporting system and a mobile application respectively through interpretability improvement and usability improvement,and obtains good results.In summary,this paper proposes a processing method for various types of medical image data for the nature of convolutional neural networks,and conducts in-depth analysis and modeling practice for many typical scenarios.The factors that need to be considered when constructing a specific model using convolutional neural networks are pointed out,and the common process of model construction under different image data structures is explored.The model construction and improvement methods proposed in this paper have strong versatility and have positive guiding significance for future medical image modeling work.