Enhancing intra-aural disease classification with attention-based deep learning models

作     者：Demircan, Furkancan Ekinci, Murat Cömert, Zafer 

作者机构：Software Engineering Faculty of Engineering and Natural Sciences Samsun University 19 Mayıs Samsun Turkey Computer Engineering Faculty of Engineering Karadeniz Technical University Trabzon Ortahisar Turkey Department of Technical Sciences of the Western Caspian University Baku Azerbaijan 

出 版 物：《Neural Computing and Applications》 (Neural Comput. Appl.)

年 卷 期：2025年第37卷第9期

页      面：6601-6616页

核心收录：

学科分类：0710[理学-生物学] 0808[工学-电气工程] 08[工学] 0836[工学-生物工程] 0701[理学-数学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：Open access funding provided by the Scientific and Technological Research Council of T\u00FCrkiye (T\u00DCB\u0130TAK). There is no funding source for this article 

主　　题：Deep learning 

摘      要：Ear diseases are defined as pathological conditions that indicate dysfunction or abnormal function of the ear organ, which is part of the auditory system of living organisms that regulates hearing and balance functions. These diseases usually manifest as conditions that affect the internal components of the ear structure and can manifest themselves with symptoms such as hearing loss, ear pain, balance problems, and fluid accumulation in the ear. The accuracy of the diagnosis depends on expert knowledge and subjective opinion. This method is prone to human error. This study presents a novel computer-aided diagnosis system for otoscope images of ear diseases, utilizing a vision transformer-based feature extractor combined with machine learning classifiers to provide accurate second opinions for ENT specialists. For this purpose, a new model based on state-of-the-art vision transformer feature extractor and machine learning models is proposed. In the experimental study, the dataset, comprising 880 eardrum images categorized into four classes (CSOM, earwax, myringosclerosis, and normal), was split into training (70%), validation (10%), and testing (20%) subsets. Each image was preprocessed to 420 × 380 pixels to fit the input dimensions of the models. The vision transformer architecture was utilized for feature extraction, followed by classification using various machine learning algorithms including kNN, SVM, and random forest. As a result, the model using vision transformer feature extractor and k-nearest neighbors (kNN) algorithm achieved 99.00% accuracy. In this study, a deep learning-based and computer-aided diagnosis system, in other words, a computational model, was developed instead of the current human error-prone disease diagnosis method used by ear nose throat (ENT) specialists. The main purpose of the deep learning-based decision support system is to support the diagnosis process where expert knowledge is difficult to access and to provide an alternative opi