Research On Lesion Segmentation And Prediction Of Coronavirus Disease 2019 Based On CT Imaging

Coronavirus disease 2019(COVID-19)is an acute respiratory infectious disease that has spread to many countries and seriously threatened the health of people all over the world.Computed tomography(CT)has high sensitivity for the detection of COVID-19 and can be used for rapid diagnosis in severely affected areas.There is an urgent clinical need for a CT-assisted analysis tool to provide reliable diagnostic information which can relieve the diagnostic pressure caused by the large number of CT images.Based on CT image data,a multi-task learning network is proposed for COVID-19 diagnosis and quantitative assessment of COVID-19 severity.Then,to further explore the lesion segmentation task,a hybrid-feature cross fusion network for segmenting COVID-19 lesions is proposed to perform supervised learning.Finally,for the domain shift problem between data from different centers,unsupervised domain adaptation methods are introduced to enhance the generalization performance of segmentation network.The specific work of this thesis is as follows.(1)Considering the relevance of disease diagnosis task and severity assessment task in clinical practice,a multi-task learning network is proposed for COVID-19 diagnosis and quantitative assessment of COVID-19 severity.The network firstly extracts taskshared features and then designs diagnosis and assessment branches to explore taskspecific information.The diagnosis branch employs multiple cascaded convolution blocks for disease identification and generates different-scaled activation features by combining intra-block features with category weights to provide guidance information for the assessment process.The assessment branch introduces an attention embedding module for effective fusion of multiple features.The proposed method achieves 87.80%identification accuracy and 66.46% assessment multi-classification F1 score,outperforming the compared classification methods.(2)The existing methods are mainly trained based on slice data,ignoring the spatial structure information in CT volumetric data.The 3D networks can fully explore spatial information,but they suffer from high computational cost and slow convergence speed.To solve this problem,a 2D multi-scale subnet and 3D lightweight subnet are designed,and by aggregating both subnets,a hybrid-feature cross fusion network is proposed for COVID-19 lesion segmentation.The network jointly learns inter-slice and intra-slice lesion features,and the cross-fusion module is designed to perform bidirectional information interaction between the features from two dimensions,which achieves complementary advantages,and extracts richer COVID-19 lesion information.The experimental results show that the proposed method can accurately segment differentsized lesions,significantly improving the segmentation performance and obtaining a Dice score of 74.85%.(3)For inter-domain bias and intra-domain gap between data from different centers,a specific alignment network based on multi-cluster centroid features and neighbouring semantic relationship,and an adaptive alignment network based on domain prototype are proposed to improve generalization performance of the network on different datasets.The former embeds specific auxiliary alignment information of multiple cluster centers from source domain into target domain samples,and makes them automatically aligned to source samples with larger similarity.Then the adversarial learning strategy is employed on the neighboring relationship maps of segmentation results to align the structure information,reducing the domain bias.The latter utilizes a single prototype from source domain combined with confidence probability map of discriminator to adaptively provide domain-level auxiliary alignment information for the target samples,which ensures the compactness of samples from source and target domains.Furthermore,a probabilistic updating strategy is designed to dynamically update the prototype to improve the discriminability of prototype.The experimental results show that the two proposed methods achieve better results than the existing methods in all three datasets and the adaptive alignment network has better performance.