Research On Medical Image Recognition Technology Based On Convolutional Neural Network

In recent years,computer aided diagnosis(CAD)has become one of the main research topics in medical imaging and radiology.Since the Annual meeting of the Radiological Society of North America(RSNA)in Chicago,the number of papers on CAD-related topics has increased substantially each year.Many different types of CAD schemes are being developed for detecting and/or describing a variety of lesions in medical imaging,including conventional projection X-Ray,computed tomography(CT),Magnetic resonance imaging(MRI),and ultrasound.Organs currently under CAD study include the chest,colon,brain,liver,kidneys,and vascular and skeletal systems.The basic concept of CAD is to provide computer output as a "second opinion" to help radiologists read images.Therefore,in order to develop a successful CAD program,it is necessary not only to develop computer algorithms,but also to study how useful computer output is for radiologists’ diagnosis,how to quantify the benefits of computer output for radiologists,and how to maximize the impact of computer output on diagnosis.Therefore,large-scale observer performance studies for radiologists using reliable methods,such as receiver operating characteristics analysis,are as important as the development of computer algorithms in the CAD field.Therefore,CAD research and development requires a team effort of researchers with different backgrounds,such as physicists,radiologists,computer scientists,engineers,psychologists and statisticians.COVID-19 is a novel coronavirus,a global outbreak that began in late 2019.With the rapid spread of the epidemic in many countries,most countries did not form effective preventive measures in time.At the same time,countries are facing a severe shortage of health resources,resulting in a surge in infections and deaths,especially in poor regions and countries.Due to the high cost of the RNA test(RT-PCR)used,the long time of the test,and the subjective judgment of medical personnel,it is easy to lead to misjudgment.These problems prompted us to develop a deep learning model to help medical personnel detect COVID-19 cases using medical images.Considering the above factors in different regions of the world,I propose the following two artificial intelligence methods for automatic recognition of COVID-19 medical images based on deep learning.(1)MLES-Net,a deep learning method for COVID-19 lung X-Ray image recognition.Considering the characteristics of high inter-class similarity and low intra-class variability of X-Ray images of COVID-19 patients,we designed a multi-level Enhanced Sense module(MLES)specifically.Based on this module,a new convolutional neural Network(MLES-Net)was proposed to classify and recognize LUNG X-Ray images of COVID-19.The first type of method is used to evaluate MLES-Net based on two open source datasets.Experimental results show that MLES-Net56-GAPFC achieves the best overall accuracy(95.27%)and recognition rate(100%)for COVID-19 categories.Compared with other methods of the same type,the MLES-Net56-GAPFC of this method also performs extremely well in model training speed,model parameter size and detection result accuracy.(2)RSP-Net,a deep learning method for COVID-19 lung CT image recognition.The patient’s CT images showed multiple small plaques and interstitial changes in the lungs,which were more obvious in the periphery of the lungs,and gradually developed multiple ground glass opacity/lung ground glass opacity(GGO)and infiltrating shadows.According to the characteristics of chest CT images of patients,we proposed a new convolutional neural Network(RSP-Net)to classify lung X-Ray images of COVID-19 by introducing the dual attention mechanism.On the open source dataset,the experimental results show that RSP-Net(parallel)has an accuracy of 94.55%.Through the analysis of comparative experimental results and comparison with other methods,the model in this paper performs well in the classification of CT images of covid-19 patients.At the same time,in order to improve the recognition efficiency,the Network model of the two methods proposed in this paper adopts three classifiers,namely,one-layer full-connection layer(FC),GAP module and GAPFC module.According to the comparison of parameter number and computation amount,the two Network models with GAPFC as the classifier have the best comprehensive performance in all aspects.