Classification of Ear Diseases from Otoscopic Images Using DenseNet121 and Grow-and-Learn Network

Diagnosis of ear diseases from otoscopy images is of critical importance to initiate the treatment process accurately and quickly. In this study, a 4-class classification problem is solved using Ear Imagery Database containing images labeled chronic otitis media, earwax plug, myringosclerosis and normal. InceptionV3, Xception, ResNet50, DenseNet121, VGG19 and MobileNet models were used to extract features from images and Grow-and-Learn (GAL) network was used for classification. In GAL network, the number of nodes is automatically adjusted and updated during the training process when needed. The network expands as it learns new information and shrinks as it starts to forget. GAL has a structure similar to KNN (K-Nearest Neighbor) algorithm in the classification phase, but it is a network that can perform higher performance classification with fewer nodes. In this way, it can be used as an alternative to KNN and the classification process can be performed in a shorter time. With this method, ear images were classified into different disease states and the performance of the proposed model was evaluated. Experimental results show that the best performance is achieved with DenseNet121+GAL with a maximum of 98.86% and an average of 97.73%. DenseNet121+KNN achieves a 96.21% success. Experiments show that GAL provides faster and more accurate classification using fewer nodes. This approach can help experts diagnose diseases quickly and accurately, thus providing more effective and timely intervention.