Research On Segmentation Method Of COVID-19 Lesion Region In Pulmonary CT Images Based On Dual Encoders

Coronavirus Disease 2019(COVID-19)is an acute respiratory infection caused by the SARS-Co V-2 virus,and computed tomography(CT)images have become an important tool in the diagnosis and assessment of patients.Automatic segmentation of COVID-19 lesion region from CT images will help improve diagnostic efficiency.At present,there are still problems in related studies,such as a lot of noise in the dataset,complex characteristics of COVID-19 lesions,and limited availability of annotated datasets.Based on the above problems,the following researches were conducted in this thesis:(1)Aiming at the problem of a large amount of noise in CT images,this thesis designs a lung region cropping algorithm used in CT images for data preprocessing.The algorithm is designed based on the morphological method,which uses binarization,maximum connected region extraction,binary image thinning,hole filling and opening operation methods to generate a binary image of the region of interest.Finally,the original CT image is cropped according to the circumscribed rectangular frame of the maximum connected region in the binary image to obtain the region of interest.Experimental results show that preprocessing with the lung region cropping algorithm can effectively improve the segmentation performance of the model.(2)In view of the feature complexity and size variety of COVID-19 in pulmonary CT images,this thesis designs a multi-scale lesion region segmentation model incorporating Transformer based on the encoder and decoder structure of U-Net.Two encoders are included in the model: a multi-scale feature enhanced encoder and a Transformer-based encoder.The multi-scale feature enhanced encoder achieves multi-scale feature extraction of the COVID-19 lesion region within the layer by employing atrous convolution.The Transformer-based encoder is mainly used for global feature extraction to compensate for the shortcomings of the multi-scale feature enhanced encoder.For an input CT image,two encoders run in parallel and perform feature encoding separately.A feature fusion module is designed in the decoder section to fuse the feature maps from the two encoders.This module enables pixel-level correspondence between the features of the two encoders and makes full use of the local and global feature information to segment the COVID-19 lesion region from the CT image.Experimental results show that the proposed model in this thesis achieves better results on both Dice and Io U evaluation metrics on the binary segmentation datasets.Better results were also achieved on the m Dice and m Io U evaluation metrics in the multi-class segmentation task.(3)A semi-supervised training strategy based on uncertainty awareness is designed for the problem of limited datasets for the segmentation task of COVID-19 lesion regions.The training method is based on mean teacher model and introduces estimates of aleatoric uncertainty and epistemic uncertainty.The student model is allowed to learn from the more reliable part of the prediction results of the teacher model.The parameters of the teacher model are updated by an exponential moving average of the student model,leading to improved segmentation performance of the teacher model.The training strategy is experimentally validated to achieve segmentation results comparable to those of the fully supervised training model using limited data.The method in this thesis can achieve accurate segmentation of COVID-19 lesion regions in CT images based on limited training data with annotation.It supports the monitoring of COVID-19 patients’ conditions during clinical treatment and can better help doctors to rapidly make treatment decisions.