Nasal And Paranasal Sinus Tumor Image Segmentation Based On Convolutional Neural Networ

Nasal cavity and sinus tumor is a malignant disease with a high incidence.It is often diagnosed at a late stage and difficult to treat,which seriously threatens the life and health of patients.Therefore,early detection and timely treatment can help improve the cure rate and prognosis of patients with nasal cavity and sinus tumor.Early examination and diagnosis often use imaging technology to scan patients,the purpose is to show the specific size and outline of the tumor,so that doctors can determine the later treatment method according to the degree of disease.Often,doctors have to make manual measurements based on the lesion area of the nasal cavity and sinus tumor on the picture.However,due to the large changes in tumor size and shape,it is difficult to accurately measure,and it takes a lot of time.And only through the subjective judgment of the doctor,it is difficult to detect some small features of the disease,and the resulting diagnosis results are often incomplete,making it difficult to achieve accurate diagnosis.Aiming at various problems at present,the deep learning algorithm is used to realize the computer-aided diagnosis of nasal cavity and sinus tumor.The two convolutional neural network frameworks based on deep learning algorithms proposed in this paper enhance the feature learning ability of the network to deal with the challenges of feature learning for nasal cavity and sinus tumor of different shapes and sizes.At the same time,due to the lack of CT image segmentation datasets required for the scientific research of nasal cavity and sinus tumor,and in order to facilitate the next research,a CT image data set of nasal cavity and sinus tumor for medical image segmentation was produced.The segmentation method of nasal cavity and sinus tumor proposed in this paper is as follows:(1)An improved DC-Unet convolutional neural network based on atrous convolution is proposed.U-Net uses skip connections to splicing the output results learned by the convolutional layers at the same level with the results of upsampling in the shrinking path to obtain a good segmentation effect.However,the algorithm still cannot fully learn the object position and detailed feature information,and cannot shield the influence of noise and build a clear object boundary.In this paper,the dilated convolution is integrated into the U-Net network,and multi-scale information is obtained by expanding the receptive field,which improves the feature extraction and learning capabilities of the model,and does not increase additional computational overhead,and improves the feature extraction capability.The experimental results finally show that the algorithm proposed in this paper can effectively improve the segmentation accuracy of nasal cavity and sinus tumor,and the highest segmentation accuracy can reach 84.12%.(2)A segmentation algorithm of nasal cavity and sinus tumor based on MD-Unet is proposed.After experimental verification,the improved U-Net segmentation algorithm has the problems of insensitivity to the learning of small target samples and splitting the local and overall consistency.First of all,by constructing a multi-scale framework,this paper extracts the features of different scales of pictures,fuses the semantic information features of large-scale images and the geometric detail information features of small-scale images,learns objects and image details of different scales,and preserves the integrity of the picture.Secondly,the deformable convolutional network is integrated,and the sampling points of its convolution kernel can follow the ability of the target shape to freely transform and learn,so as to obtain the receptive field of free transformation,and further learn the detailed characteristics of the object.Finally,Tversky is used as the loss function to address the data imbalance problem.The implementation results show that the segmentation accuracy of the MD-Unet algorithm on the nasal cavity and sinus tumor dataset can reach 84.27%.The final experimental results show that the two improved algorithms proposed in this paper perform better in the segmentation of nasal cavity and sinus tumor datasets,improve the segmentation accuracy,and help doctors make accurate diagnosis.