Detecting external and middle ear diseases using deep learning algorithms

Ear diseases are easily treatable with early detection and appropriate medical care. However, the small number of experts and the relatively low diagnostic accuracy require a new diagnostic strategy in which deep learning can play an important role. In this thesis, deep learning models are used to detect ear diseases based on a large number of autoendoscopic images obtained in the clinical setting and the success rates of these models are shown. In this thesis, a data set with 10 disease classes (Normal, Adesiv, Akut Otit, Eksternal, Miringoskleroz, Opere, Perfore, Seröz, Serümen, Süpüratif Kronik Otit) covering ear diseases was used. In this data set, there are a total of 2,125 autoendoscopic images, of which 1049 are healthy (normal) and 1076 are diseased. In addition, various pre-processes and data augmentation processes were applied to this data set. Convolutional Neural Networks (CNN) method was used to classify the diseases in this data set. Pre-trained models EfficientNetB2, EfficientNetB3, InceptionV3, Resnet50, Resnet50V2, ResNet101, ResNet101V2, VGG-16, Xception and InceptionResNetV2 were used to perform the classification process, and two hidden layers consisting of 2048 and 1024 neurons were added to the fully connected layers of these models. In this thesis, experiments were carried out with different optimizer algorithms and different epoch-batch size values. As a result of the optimizer experiments, the Adam optimizer gave the best results and was used in the epoch-batch size experiments. EfficientNetB2 and ResNet50 models gave the best results in epoch-batch size experiments, with an accuracy of 96% in 40 epoch-32 batch size studies. Adesiv and Opere diseases were the diseases best recognized by CNN models.