Research On Nasal CT Image Segmentation Based On Depth Learning

Multiple scientific studies have found that the incidence of sinusitis is significantly associated with abnormal deviation of the nasal septum,and it is one of the most common chronic diseases in ENT.Deviated nasal septum is a common anatomical variation of nasal cavity and an important cause of sinusitis.Therefore,the precise segmentation of nasal septum plays a very important role in assisting doctors in clinical diagnosis and treatment.With the development of digital medical equipment,doctors increasingly rely on medical imaging for disease diagnosis.CT imaging is a common method for nasal examination and diagnosis.Due to the different shapes of nasal septum,blurred boundaries,and often accompanied by mucus,there is a large subjectivity in artificial diagnosis,which cannot be accurately identified.In recent years,more and more medical imaging fields have introduced deep learning technology,and achieved very good results.It is often used to detect and segment organs and lesion areas.Accurate segmentation of medical image is the basis of disease analysis,it will be based on specific medical image tasks to be segmented areas,according to some similar features of the image segmentation results.In order to accurately segment the nasal septum region during nasal surgery,the SAMU-Net model is proposed in this paper to segment the nasal septum region.Compared with the classical medical image segmentation method,the method proposed in this paper has higher segmentation accuracy and better effect.The research work of this paper is as follows:(1)This paper collected the data set of snoring patients in a Class A hospital of Dalian,and then preprocessed the data set.Sections were performed along the sagittal,transverse and coronal axes.According to the position distribution of nasal septum,the most obvious coronal sections in the nasal septum region were selected from the 3D data to label the nasal septum region and construct the data set.Through data augmentation and expansion,the samples become richer and more diverse.In order to make the model have better generalization,5 folds cross validation is used in training.(2)Aimed at the noise interference in CT images,the different shapes of nasal septum in nasal cavity,and the fuzzy boundary problem,the SAMU-Net model was proposed.SAM attention mechanism module was added after the second convolution of each convolution layer to suppress the expression of unrelated areas and enhance the ability of network to extract information.In view of the imbalance between positive samples(nasal septum tissue is too small)and negative samples(background is too large)in the data,small features are not obvious,and it is difficult to segment,an improved focus loss function is introduced,and the loss function is expanded into a polynomial combination through Taylor.By adjusting the first polynomial coefficient,the model pays more attention to the learning of positive sample area.Experiments show that adding SAM attention mechanism and improving the focus loss function can make the network more robust and significantly improve the segmentation performance.(3)Design and develop a system interface for nasal CT image segmentation visualization.By loading the image to be segmented onto the system interface and running the system,you can observe the visualization results after segmentation.The interface is clear and smooth,and the operation is simple.