The Research On Segmentation And Classification Method Of COVID-19 CT Image Based On Neural Network

The global outbreak of novel coronavirus pneumonia(Corona Virus Disease2019,COVID-19)since December 2019 has had a serious impact on daily life.Computed Tomography(CT)results are key bases for diagnosing COVID-19.Doctors first segment the lesion area and then diagnose the lesion’s interior features in the diagnostic process.Lesion segmentation depicts the focal regions in CT images,providing a basis for assessment and quantification;disease classification can quickly diagnose cases with high accuracy,aiding doctors in diagnosis.The lesion areas in the chest CT images of COVID-19 patients show ground glass density shadow,solid white like pulmonary solids,Etc.The scenes are complex,and the blurred lesions pose a great challenge for lesion segmentation work.Moreover,the similarity of imaging features between COVID-19 lesions and other pneumonia poses difficulties for disease classification.In addition,the results provided by classification networks alone are difficult to support physicians in making a credible diagnosis.The lesion segmentation results can improve the performance of the classification network;the disease classification task can also enhance the segmentation network’s focus and strengthen the segmentation of the edge part.Multi-task learning approach can mutually enhance both the performance.The following four are accomplished in this paper to optimize the COVID-19 CT image lesion segmentation and disease classification tasks.(1)Annotation of COVID-19 data and dataset construction: This paper collaborated with Shanghai Sixth People’s Hospital and Shanghai Ninth People’s Hospital to collect 272 chest CT data from 2020 to 2021.A lung CT image segmentation dataset including 100 COVID-19 patients and a multi-task classification segmentation CT image dataset of 272 different pneumonia patients were constructed for subsequent experimental validation.(2)COVID-19 lesion segmentation method based on 3D deformable convolution: For the problems of irregular and blurred edges of COVID-19 lesions diverse and complex imaging features,we propose DF-Net,a COVID-19 lesion segmentation method.3 Dimensional Deformable Convolution(3DDC)module is added to DF-Net to adaptively adjust the shape of the convolution kernel to extract the features of COVID-19 irregular lesions;the Focal Tversky loss function is introduced to balance the false positive and false negative voxels;the migration learning strategy is used to alleviate the problem of small COVID-19 dataset.Experiments on the constructed lung CT image segmentation dataset showed that the method achieved 83.04% Volume Dice and 79.97% Surface Dice,respectively.(3)Multi-task pneumonia classification and segmentation method based on reverse attention mechanism and iterative training strategy: For difficult classification problems of COVID-19 and other pneumonia,we propose Multi R-Ne,multi-task disease classification and lesion segmentation methods to assist COVID-19 diagnosis.Multi R-Net consists of two sub-networks: a U-shaped subnet for 3D lesion segmentation;and a residual subnet for disease classification.Feature fusion between the two subnets is performed through a reverse attention mechanism and an iterative training strategy.The method achieves a Precision of 94.01% and an F1-score of 93.66%,which outperforms existing comparison models.Visual activation graph analysis demonstrates the interpretability of Multi R-Net results.(4)Distributed training method for COVID-19 classification and segmentation network: For the problem that a large amount of CT scans and a large number of network model parameters lead to long network training time,we propose a Tensorflow-based multi-GPU distributed training method,using a data-parallel training strategy and multiple communication architectures to optimize the training speed.The experimental results show that the method accelerates the network training by 3.66 times on 4 GPUs without any significant loss of metrics.