Construction Of The Fiber Electronic Laryngoscope Image Recognition Model On The Basic Of Deep Learning And Its Application In The Diagnosis Of Laryngeal Precancerous Lesions

Part ?: Construction and evaluation of the fiber electronic laryngoscope image recognition model on the basic of deep learningObjective: Laryngeal malignancies are common malignancies of the head and neck.Before the normal laryngeal epithelial tissue becomes laryngeal cancer,it usually passes through the stage of laryngeal precancerous lesions.In this study,the purpose is to build the fiber electronic laryngoscope image recognition model on the basic of deep learning,to put the fiber electronic laryngoscope image in the different stages of laryngeal precancerous lesions and early laryngeal cancer into the model,and to evaluate its diagnostic efficacy.Method: A total of 6515 cases of laryngoscope images of laryngeal precancerous lesions and early laryngeal and normal laryngoscope images were collected from the ENT laryngoscope room of Changhai Hospital.Each image had a clear corresponding pathological result.The processed images were divided into the training sets(mild to moderate atypical hyperplasia in 2550 cases,severe atypical hyperplasia in 1944 cases,early laryngeal cancer in 1571 cases)and the testing sets(mild to moderate atypical hyperplasia in 150 cases,severe atypical hyperplasia in 150 cases,early laryngeal cancer in 150 cases).The laryngoscope images of the training sets were pre-processed.The training was performed by using three models of VGG-19,Res Net50 and Inception-V3,and after the above three models were transformed,images the training sets were used for model training.After model training was complete,images the testing sets were used to validate the model.Results: AUC of Res Net50 model on mild to moderate atypical hyperplasia,severe atypical hyperplasia,and early laryngeal cancer are 0.873(95%CI,0.774-0.971),0.879(95%CI,0.794-0.964),0.857(95%CI,0.743-0.971).Among the three models,the performance is the best,the model is stable,and the results are reliable.Conclusion: The fiber electronic laryngoscope image recognition model constructed on the basic of Res Net50 model can make a more accurate diagnosis of laryngeal precancerous lesions and early laryngeal cancer,and shows good diagnostic results.Part ?:The applying of fiber electronic laryngoscope image recognition model on the basic of deep learning for the diagnosis of laryngeal precancerous lesionsObjective: The morphological manifestations of precancerous lesions and early laryngeal cancer with different levels by fiber electronic laryngoscopes are similar.It is difficult to distinguish them with the naked eye.The diagnosis is mainly on the basic of the clinical experience of endoscopic physicians.The purpose of this study is to explore the application value of fiber electronic laryngoscope image recognition model on the basic of deep learning in assisting clinicians in the diagnosis of laryngeal precancerous lesions and early laryngeal cancer.Methods: Sixteen clinicians from the Department of Otolaryngology Head and Neck Surgery in our hospital were recruited.Among them,there were 8 attending physicians with more than 5 years of experience in electronic laryngoscope operation and diagnosis,and there were 8 attending physicians with less than 3 years of experience in electronic laryngoscope operation and diagnosis.First,150 fiber electronic laryngoscopes images(50 mild to moderate atypical hyperplasia images,50 severe atypical hyperplasia images,and 50 early laryngeal cancer images)were collected,and the physician was asked to diagnose each image separately.After 150 electronic laryngoscope images were put into the electronic laryngoscope image recognition model for diagnosis,the image containing the diagnosis results of the model was provided to the physician,and the physician was asked to give a diagnosis to 150 images again.Two diagnosis results of each physician were recorded separately,diagnostic accuracy,sensitivity,and specificity of each physician before and after model assistance were calculated,and t test was used to analyze the data.Results: Before assistance of the electronic laryngoscope image recognition model,the attending physician with more than 5 years of experience in the operation and diagnosis of electronic laryngoscopes had an average diagnosis accuracy rate of 81.75,for the three groups of diseases,and the attending physician with less than 3 years of experience in the operation and diagnosis of electronic laryngoscopes had an average diagnosis accuracy rate of 72.25%,for the three groups of diseases.After model assistance,the attending physician with more than 5 years of experience in electronic laryngoscope operation and diagnosis had an average diagnosis accuracy rate of 85%,for the three groups of diseases.The attending physician with less than 3 years of experience in the operation and diagnosis of electronic laryngoscopes had an average diagnosis accuracy rate of 82.5,for the three groups of diseases.The accuracy rate,of the two groups of physicians in the diagnosis of the disease were improved after model assistance.Conclusion: In this study,the constructed fiber electronic laryngoscope image recognition system on the basic of deep learning has certain value to the attending physicians of different years in the diagnosis of laryngeal precancerous lesions and early laryngeal cancer.To a certain extent,it can improve the diagnostic efficacy of clinicians for this type of disease.