| Monitoring the course of pneumonia caused by novel coronavirus(COVID-19)is an effective means to control the rate of severe conversion of neoconiosis.Ultrasound instruments are convenient and quick,and are widely used based on real-time acquisition of lung ultrasound images.When the ultrasound signs of patients with COVID-19 develop from B-line status to white-lung status,it indicates that the disease deteriorates to severe disease,so it is especially important to screen and determine the B-line and white-lung signs of patients in a timely manner.However,clinical discrimination of B-line and white lung is highly dependent on physicians’ experience,which is highly prone to omission and misdiagnosis.In recent years,tasks such as detection and classification have been trained with algorithmic models that perform well in terms of robustness and generalization ability with the help of deep learning algorithms.In this paper,we design and implement an intelligent classification system for B-line and white lung,two representations of COVID-19 in lung ultrasound images based on deep learning,and carry out research in three aspects,including pleural line detection algorithm,classification algorithm between B-line and white lung,and application development,as follows.1.Since the pleural line on the image has a positional association with the B-line and white lung region,a pleural line detection algorithm is proposed in this paper to indirectly obtain relatively accurate B-line and white lung lesion locations and filter redundant information.The algorithm framework is based on the YOLOv5 network,using the lightweight network Goog Le Net as the feature extraction network in the Backbone layer,and using the Bi FPN feature pyramid structure to optimize the Head layer in the framework,while using the EIo U loss function with the aspect ratio as the focus of attention,and finally using CA the attention mechanism to obtain the spatial location dependency.The performance of the improved algorithm achieves precision of 96.9% and recall of 92.4%.2.After the pleural line detection process to obtain the relatively accurate B-line and white lung region,the classification algorithm between B-line and white lung is proposed.The algorithm is based on the classical convolutional neural network Res Net,with the addition of SE attention mechanism and Non-local self-attention module in terms of network structure.This modification captures channel-level and remotely dependent feature representations between any two locations,allowing the model to mine potential information from both channel and spatial perspectives.The improved algorithm improves the AUC by 4% compared to the base algorithm.3.In this paper,we design the development system with B/S structure aiming to deploy the usage model online.The front-end design of the system is based on the combination of Vue,Element-UI and Axios to realize multiple page logics such as login,registration,upload,etc.The back-end uses Spring Boot as the main framework to deploy the above detection and classification algorithms through Flask,and combines the database My SQL for accessing business information operations to realize the combination of deep learning algorithms and Web engineering to enhance the system usability. |
