Segmentation Of COVID-19 Lesions In Lung CT Images Based On PAID-Unet++ Model

The current New Coronavirus epidemic is now a global concern,and CT scan image analysis of the lungs is one of the most important methods for diagnosing and detecting New Coronavirus(COVID-19)infection.In recent years,deep learning has given new prospects for the segmentation of COVID-19 lesions in lung CT images,which is an important computer-aided diagnostic technique to automatically segment COVID-19 infected regions in lung CT images.This paper uses the UNet++ model with relatively good segmentation accuracy as the backbone network for subsequent research,and on this basis incorporates various mechanism modules that can optimise the segmentation effect and performance,in the process of forming the final PAID-Unet++model,the main research work carried out in this paper is as follows:(1)In order to fully utilize the attention resources in the model to the key pixel locations in the infected region,this paper forms a new mixed attention mechanism module by concatenating spatial attention with channel attention to enhance the availability of the model to extract features.In order to solve the problem of low accuracy and sensitivity of model segmentation brought about by the high imbalance between positive and negative samples in the COVID-19 dataset,yet the specificity index is falsely high,this paper uses the Focal Tversky loss function to optimize the training process of the model and reasonably assign the penalty weights of the loss function.(2)To address the problem of gradient disappearance in the training process of deeper models,this paper designs a progressive encoder and decoder by replacing the forward convolution operation only in the original codec with a residual unit structure with a recurrent convolution layer respectively,to enhance the representational capability of the model.Since the infected regions are widely and irregularly distributed on CT scan images,this paper adds a convolutional perceptual field for this purpose and uses the dilated convolution mechanism in order to extract the long-distance dependencies in the feature pictures,and on this basis,the dilated convolution mechanism with different perceptual fields composed of different dilational rates is designed in turn according to the different input feature picture resolutions required by the decoder in each branch of the model,in order to fit the different layers of the The decoder’s different feature map acceptance fields are designed to fit the needs of each level.(3)Due to the fact that the model introduces more modules of various mechanisms and its own structure is more complex,the number of parameters to be computed in the experimental process of the model is larger.Based on this,this paper uses the depth-supervised approach for model pruning,although the segmentation accuracy will be reduced to a certain extent,the number of parameters computed in the lightweight model will be significantly reduced from 117.1M to 71.9M,which improves the computational efficiency.To verify the performance of the PAID-Unet++ network,the Dice similarity coefficient,Miou,sensitivity and specificity evaluation metrics were used to quantify the model segmentation results.The qualitative and quantitative experimental results on the COVID-19 CT scan dataset showed that the Dice similarity coefficient and the average intersection ratio improved by 2.85% and 2.7%,respectively,compared with the backbone model Unet++,although the specificity index decreased by 1.16%.At the same time,it also penalizes the excessive false negatives in the segmentation results of the model,and thus increases the sensitivity of the evaluation indexes focusing on the detection of true positives and false negatives by 3.05%,and achieves the accuracy values of DSC,Miou,Sensitivity,Specificity in each evaluation index respectively.68.52%,60.11%,74.19%,and 90.57%.Through comparison experiments with other segmentation models,it can be seen that the PAID-Unet++ model proposed in this paper outperforms most advanced segmentation models in terms of segmentation effect,and can accurately segment the infected region of COVID-19 CT scan images,which has good prospects for clinical diagnosis.