Auxiliary Diagnosis Method And Application Realization Of Nasal Lesions Based On Deep Learning

Affected by environmental factors and genetic factors such as the increase in haze weather,the incidence of nose-related diseases is increasing year by year.Common nasal diseases include nasal polyps,fungal sinusitis,and nasal cavity tumors.At present,nasal endoscopy is a routine method for examination and surgical treatment of these diseases.However,the nasal lesions are difficult to distinguish and concealed,which leads to a high rate of misdiagnosis.Now application of deep learning techniques in medical image processing is wide,and how to apply deep learning to assisted diagnosis of nasal lesions still needs in-depth research.In this paper,we collected the videos of nasal endoscopic surgery of 202 patients,intercepted the video frames,selected 3041pictures to mark the lesion area,established a nasal endoscopic image dataset,and studied the auxiliary diagnosis method of nasal lesion area with deep learning.The main work includes:In order to recognize the lesion area with high accuracy,the DAC and RMP modules are embedded in the U-Net for the recognition problem of the lesion area with multi-scale nasal endoscopic image besides using cascaded atrous convolution and pooling.These operations to intensive extraction of image features of different sizes to realize the recognition of lesions.Aiming at the problem that the color of the lesion area is similar to the normal area and it is difficult to distinguish,the attention mechanism including Transformer is introduced in the U-Net.Through multiple iterations of training,the network’s attention is focused on the lesion area.And use Res2Net as the backbone feature extraction network,the Loss function takes the weighted average of BCE Loss and Dice Loss.The experimental results show that compared with Deep Snake,U-Net,Att-U-Net and CE-Net,the Io U and Dice Scores of Art Ce Net increase to 87.7%and 92.7%respectively.In order to realize real-time tracking of lesions in nasal endoscopy video,this paper studies the offline tracking algorithms TLD and KCF based on detection and the online tracking algorithm Siam Mask based on deep learning.Aiming at the shortcomings of insufficient spatial information of Siam Mask,a multi-scale feature fusion module is proposed.Based on the speed and accuracy of fusing features of different scales,the optimal configuration of Siam Mask-U_2,3,4 is given.It’s recall rate reaches 87.37%and speed is 32.967fps.Finally,use the recognition results of Art Ce Net to initialize Siam Mask-U_2,3,4 to realize the complete intelligent recognition and tracking diagnosis of the lesion area in the video of the nasal endoscope.Finally,Art Ce Net and Siam Mask-U_2,3,4 are deployed on the B/S framework to build a nasal endoscopy-assisted diagnosis system.Experiments show that the system initially realizes the auxiliary intelligent diagnosis of nasal common disease and lays the foundation for telemedicine.