| Purpose and significanceLaryngeal cancer is one of the most common malignant tumors in the head and neck,ranking the third in the incidence of head and neck malignant tumors,and the 21st in the general tumor.Surgical treatment has always been the main means of laryngeal cancer treatment According to relevant studies,the incidence of laryngeal cancer recurrence after surgery is as high as 16%-29%,and recurrence will greatly reduce the five-year overall survival rate and overall survival rate.Some patients fail to find recurrent lesions in time,leading to disease progression,laryngeal preservation therapy,permanent loss of laryngeal function.Therefore,for patients with recurrent laryngeal cancer after surgery,the earlier the lesion is detected,the laryngeal function can be preserved as far as possible,so as to guarantee the quality of life and improve the prognosis.Electronic fiber laryngoscopy is the most intuitive and widely used examination of larynx,which can detect relatively hidden lesions in the early stage.However,the diagnosis of electronic fiber laryngoscopy usually depends on the experience of endoscopists,and the diagnostic level of endoscopists in clinical practice is uneven,and there are certain misdiagnosis and missed diagnosis.Therefore,improving the diagnostic methods and improving the diagnostic accuracy of postoperative recurrence of laryngeal cancer is of great significance for improving the prognosis of patients and safeguarding the quality of life.In recent years,research on laryngoscope image recognition model assisted in laryngeal cancer diagnosis based on artificial intelligence technology has appeared at home and abroad and achieved good results.However,there has been no research on artificial intelligence-based assisted diagnosis of laryngeal cancer recurrence.In this study,an electronic laryngoscope image recognition model based on artificial intelligence technology will be constructed for the first time and used to detect whether patients with laryngeal cancer relapse after surgery.Materials and methodsFirstly,220 patients with laryngeal cancer confirmed by pathology admitted to the otolaryngology Department of Guangdong Provincial People’s Hospital from January 1,2010 to December 3 1,2020 were retrospectively collected.All the 211 patients underwent surgical treatment in our hospital,among them,29 patients had postoperative recurrence,and 191 patients had no recurrence.A total of 7132 electronic fiber laryngoscopy images of 220 patients were collected and soiled out.546 laryngoscopy images of the patients undergoing surgery were selected as experimental data,covering different time points before and after surgery.Colabeler software was used to manually delineate laryngeal masses in 546 images to extract quantitative image features.The maximum correlation minimum redundancy algorithm,LASSO algorithm and Logistics regression model were used to screen out key feature sets.Then,five different types of traditional machine learning models(Logistics regression,artificial neural network,K-nearest neighbor,random forest and support vector machine)were constructed according to the feature set to predict the postoperative recurrence of laryngeal cancer,and the performance of these five models was compared.In addition,in order to compare the potential and performance of deep learning algorithms in laryngoscope image analysis,we also built a deep learning model of unsupervised learning--convolutional autoencoder to train and enhance 630 preoperative laryngoscope images of 220 patients.Then,after extracting the features of the enhanced images,five different types of traditional machine learning models are constructed for further analysis and comparison.ResultsBased on the candidate feature set of preoperative laryngoscopy images and the feature set of preoperative images enhanced by deep learning network,we constructed five prediction models respectively and made further comparison.The experimental results showed that the Logistics regression model achieved the highest cross-validation accuracy in both groups(preoperative image group:0.836,enhanced preoperative image group:0.868),and high stability.The sensitivities and specificities of logistic regression models for predicting postoperative recurrence of laryngeal cancer in the preoperative image group and the enhanced preoperative image group were 0.727 and 0.636,respectively.According to the candidate feature sets of laryngoscopy images at different time points after surgery,we also consstructed and compared five different types of machine learning models.By leaving a cross-validation,we found that the Logistics regression model had the best performance in the 0-6 months after surgery group and the 2-5 years after surgery group.The accuracy of cross-validation was 0.768 and 0.889,the sensitivity was 0.857 and 0.971,and the specificity was 0.619 and 0.636.In the 6-12 months and 1-2 years postoperative groups,support vector machine model performed best,with cross-validation accuracy of 0.929 and 0.806,sensitivity of 1.000 and 0.741,and specificity of 0.667 and 0.850,respectively.ConclusionThe electronic laryngoscope image recognition model based on artificial intelligence technology has a very good prediction ability for the auxiliary diagnosis of laryngeal cancer recurrence after surgery,and has a potential application prospect in laryngoscope review after surgery.The performance of the prediction model based on the enhanced deep learning network image is better than that based on the original image only. |
PDF Full Text Request