Hybrid CNN-ViT Ensemble for Enhanced Ear Disease Classification

Diagnosing ear diseases accurately is a key challenge due to considerable difference between the visual symptoms for each disease. To improve classification performance, this study presents a so-called hybrid ensemble learning model, which combines two CNNs (for RGB images and for grayscale images) and the ViT. The combination introduces complementary strengths of each architecture by learning both localized spatial features and long-range dependencies. The best individual performance came for the grayscale CNN, reporting an accuracy of 98.75 %, followed by the ViT and RGB CNN with 96.15 % and 94.20 %, respectively. Utilizing ensemble averaging, the integrated model achieved a high accuracy of 96.42 %, greatly enhancing diagnostic consistency. Performance analysis and visual validation showed stability and precision of the ensemble. This multimodality pipeline shows great potential for real-world clinical implementation and presents a scalable, AI-based solution for the accurate and automated classification of ear diseases using otoendoscopic images.