| Objective:Otitis media in children presents a high incidence and is one of the most common diseases for which antibiotics are prescribed.Otoscopy is a routine examination in the clinic.Accurate diagnosis of otitis media can bring early detection,better treatment,and proper follow-up of otitis media.To achieve automated diagnosis and improve the diagnostic accuracy,based on deep learning we developed classification models using convolutional neural networks(CNNs)for otoscopic images of children with otitis media.To create a newartificial intelligence(AI)strategy for screening and monitoring children with otitis media,we further compared the diagnostic performance of the CNNs in images captured by smartphone with WI-FI connected otoscope.Materials and Methods:The endoscopic images of children diagnosed as acute otitis media(AOM),otitis media with effusion(OME),and normal ear were retrospectively collected from the otolaryngology endoscopic database of Shenzhen Children’s Hospital.The obtained images were manually divided into three types according to the criteria,and the dataset was randomly divided into a training set and a test set.Two mainstream pre-trained models,Xception and MobileNetV2were selected.With transfer learning,the last fully connected layer with an output size of 1,000 in the original network was replaced by a new fully connected layer with an output size of three types.The accuracy,sensitivity,specificity,receiver operating characteristic curve(ROC),area under the ROC curve(AUC),and class activation mapping(CAM)of the two CNNs were evaluated.The images obtained by smartphone with WI-FI connected otoscope were prospectively included according to the above three types.This new dataset with smartphone images was tested by the two CNNs.The classification and visualization performances of the two different image types with CNNswere evaluated and compared.Results:A total of 12,203 endoscopic images(from 7,053 children)were included,including three types of AOM,OME,and normal ear.A total of 10,703 images formed the training set,including 3,355(31.3%)of AOM,4,113(38.4%)of OME,and 3,235(30.2%)of normal ears.The test set included 1,500 images composed of three types,each with 500 images.After 10 repetitions of training the dataset with CNNs,the overall accuracies of Xception and MobileNetV2 for classification were 97.45%(95%CI:96.81%-97.94%)and 95.72%(95%CI:95.12%-96.16%).A total of 102 smartphone images(from 85 children)were included to form the test set,including 26 of AOM,35 of OME,and 41 of normal ears.After 10 repetitions of training on this test set using two CNNs,the overall accuracies of Xception and MobileNetV2 for classification were 90.66%(95%CI:90.21%-90.98%)and 88.56%(95%CI:87.86%-90.05%).The performance showed the good generalization ability of the model.When calculated by confusion matrix,the classification indicators of accuracy,sensitivity,specificity,and AUC with endoscopic images were all better than those with smartphone images.Although the clarity of the original smartphone images was lower than that of the endoscopic images,the CAMs showed the extracted features of these two different images were almost the same.Conclusions:We have developed deep learning algorithms for the successfully automated classification of pediatric AOM and OME with otoscopic images.With a smartphone-enabled wireless otoscope,AI may assist parents in early detection and continuous monitoring at home to decrease the visit frequencies.Especially during the COVID-19 pandemic,telemedicine with AI is of great significance in reducing crowd gathering risk and decreasing the social medical burden. |
