Research On Early Diagnosis Of Laryngeal Endoscopy Based On Deep Learning

Objective:Laryngeal Carcinoma(LC)is the most common malignant tumors in the head and neck,and ranks the second in the incidence of respiratory tumors.In recent years,despite significant advances in the treatment of laryngeal cancer,there has been no significant improvement in survival rates.Since the survival rate is closely related to the accuracy of disease detection,the early detection of precancerous lesions is an important determinant of the prognosis of laryngeal cancer.At present,electronic laryngoscopy is often used for early diagnosis in clinical practice.However,the morphological manifestations of precancerous lesions at different stages may be similar under electronic laryngoscopy,and the diagnosis mainly depends on the clinical experience of endoscopists in the department of otolaryngology.Therefore,this study aims to develop an objective and stable method to construct a Lary Mind model based on deep learning,which can be used for focal identification and disease classification of laryngeal endoscopy images,and provide valuable reference for the diagnosis of common laryngeal diseases.Methods: A total of 8000 gastroenteroscopy images in Kvasir database were selected to construct the pre-training data set.A total of 2271 general clear white light images of electronic laryngoscopy of patients undergoing electronic laryngoscopy in the laryngoscopy room of department of otolaryngology & head and neck surgery,affiliated hospital of inner mongolia medical university from February 2010 to May 2022 were retrospectively collected.Among them,611 images of squamous cell carcinoma,158 images of vocal fold leukoplakia,854 images of vocal ford polyp,202 images of vocal ford nodules,99 images of vocal ford granuloma,300 images of vocal ford inflammation,40 images of squamous cell carcinoma complicated with vocal ford leukoplakia,7 images complicated with vocal ford polyp,and3573 images of healthy people.A total of 5844 laryngeal endoscopic images were used for model training.The performance differences of different deep learning methods in endoscopic image research tasks were analyzed through ablation experiments to determine the best model and parameters.In addition,113 laryngoscopy images were collected as test data set to compare the performance of the model and otolaryngologist.SPSS 20.0 software was used for paired T-test to compare the difference in the diagnostic accuracy of doctors before and after the model.Results: All the images were verified by histopathology,and the AUC value of the electronic laryngoscope image recognition model proposed in this paper reached 0.76 in the classification results of various laryngeal diseases,which could accurately classify various laryngeal diseases.According to the diagnostic test of 113 laryngoscope pictures,the average diagnostic accuracy of this model and doctors in the junior,senior and expert groups for the above 6 common laryngeal diseases were 89.07%,72.74%,78.11% and 82.30%,respectively.Moreover,the video reading time of the electronic laryngoscope image recognition model was significantly shorter(0.07 s vs 6.37s).And before and after model assistance,the diagnostic accuracy of doctors in the low seniority group,the high seniority group and the expert group was improved,respectively(72.74% vs 81.55%;78.11% vs 85.07%;82.30% vs 88.00%).Conclusion: The Lary Mind model proposed in this paper has a strong ability to classify common laryngeal diseases in electronic laryngoscopy images,and its accuracy is significantly higher than that of doctors in the junior group,senior group and expert group.It can provide a valuable reference in clinical auxiliary diagnosis,and has a good clinical application prospect.