Development, testing and feasibility of a diagnostically independent mobile application for eardrum otoscope images based on deep learning

Ear diseases are health problems that negatively affect our daily lives. These diseases are diagnosed by experts through observation using an otoscope device and the accuracy of this diagnosis system depends on expert knowledge. This study investigates the diagnosis of ear diseases using deep learning models. Ear Imagery and Tympanic membrane datasets were used in this study. In addition, a separate dataset was created by combining these two datasets. The datasets used were divided into 70%, 10% and 20% for training, validation and testing respectively. ResNet50, InceptionV3, InceptionResNetV2 and Xception models were selected as deep learning models. In addition, hybrid models were created from DenseNet201, EfficientNetB5 and InceptionResNetV2 feature extractor models and BiLSTM and ConvBiLSTM classifier models. Within the scope of the study, zero training, learning transfer, HSV and LUV color spaces were used to investigate the effect on classification success. As a result of the study, the highest accuracy was obtained from DenseNet201-BiLSTM and ResNet50 models.