A Deep Learning-based Algorithm For Segmenting Medical Images Of Pneumonia

The new crown epidemic has now become a global pandemic,endangering the physical and mental health of the population.Nucleic acid testing is the gold standard in the process of screening for neo-crown,and CT scans of patients can provide information complementary to nucleic acid testing.However,as collecting and processing CT images of patients is labour-intensive,methods that can automatically segment areas of neocoronary pneumonia lesions are needed to alleviate the pressure of insufficient medical resources.A number of deep learning-based image segmentation methods have achieved good performance on the neocoronal lesion segmentation task,and some studies have expanded the neocoronal pneumonia dataset with unsupervised and semi-supervised methods,and these works have greatly facilitated the development of subsequent work.Based on these problems,this paper proposes a new coronal lesion region segmentation method,UDNet,for low-contrast images.The noise-added annotated images are obtained by a probabilistic diffusion model,and the features are fused with the lesion segmentation network to learn the ability to segment lesion regions from the noise,and the loss function is redesigned to improve the quality of segmented lesions.To evaluate the new crown lesion region segmentation method UDNET proposed in this paper,three evaluation metrics,Dice coefficient,MIo U and MPA,were used on the COVID-Semi Seg and COVID-19-CT-Seg datasets.By analysing the experimental results on the two datasets,the algorithm UDNet proposed in this paper improves the Dice metric by 0.82% and 1.61%,the MIo U metric by 0.56% and 1.64%,and the MPA metric by 0.89%and 1.01% over the benchmark model UNet in semantic segmentation,and these experimental results validate that the UDNet proposed in this paper has a good performance in the new crown lesion segmentation These experimental results validate that the UDNet proposed in this paper has some improvement over previous methods in the task of new crown lesion segmentation,and the effectiveness of the strategy proposed in this paper is verified by ablation experiments.