Feature Selection Using Neighborhood Component Analysis With Deep Learning-Based for Multi-Classification of Middle and External Ear Conditions

Ear diseases are common in childhood and significantly increase the probability of developing serious complications, such as speech disorders, intracranial infection, and hearing loss. Otoscopic examination is crucial for diagnosing middle and external ear diseases. Deep learning-based computer-aided systems hold great promise for the automatic evaluation of otoscopic images and the prediction of patient outcomes. In the study, an image processing-based model was developed for the multi-classification of middle and external ear diseases. We employed histogram equalization for image enhancement and then employed the bilateral filter to reduce the noise of otoscopic images. Image feature vectors were extracted using ResNet101, DenseNet201, AlexNet, and VGG19 models. The neighborhood component analysis (NCA) was employed for distinctive feature selection. Then, the performances of classification models, including bidirectional long-short-term memory (B-LSTM), convolutional neural networks (CNNs), decision tree (DT), support vector machine (SVM), and k-nearest neighbor (KNN), were compared. The B-LSTM algorithm with the NCA feature selection method reached the highest performance and promising results with 0.985 kappa statistics, 0.988 weighted-F1 score, and 98.86% accuracy. The results demonstrated that the image processing-based deep learning model can accurately and efficiently detect middle and external ear diseases from otoscopic images. Moreover, the study outperformed the known related studies.