What is it about?
Endoscopic imaging helps doctors examine the ear, nose, and throat (ENT) to detect diseases early. However, many anatomical areas look very similar, making automatic recognition difficult. Our research introduces HyMoENet, a deep learning model that combines CNNs for fine local details, Transformers for overall context, and a Mixture-of-Experts mechanism that allows the network to specialize in different visual patterns. Tested on real clinical data from Thong Nhat Hospital, Vietnam, HyMoENet achieved 97.5% accuracy, providing a strong foundation for intelligent medical diagnostic systems.
Featured Image
Photo by Katarzyna Zygnerska on Unsplash
Why is it important?
This work demonstrates how combining multiple AI techniques can significantly improve accuracy in medical image analysis, even with limited data. By effectively distinguishing between highly similar anatomical regions, HyMoENet can assist doctors in faster and more consistent diagnosis, helping reduce human error and supporting AI-assisted healthcare in developing countries. It also sets a new benchmark for ENT image classification, showing how hybrid and expert-based architectures can advance clinical AI research.
Perspectives
As the lead author, I was motivated by the challenge of building an AI system that could truly assist clinicians in real hospital environments. Developing HyMoENet allowed our team to explore how hybrid CNN-Transformer models and Mixture-of-Experts architectures can bring practical improvements in diagnostic accuracy. We hope this work inspires further collaboration between AI researchers and medical professionals to build more reliable, human-centered diagnostic tools.
Trong Nhan Nguyen
FPT University
Read the Original
This page is a summary of: HyMoENet: Mixture-of-Experts Enhanced CNN-Transformer Hybrid Framework for Classifying Anatomical Sites in Endoscopic ENT Images, October 2025, ACM (Association for Computing Machinery),
DOI: 10.1145/3746027.3762092.
You can read the full text:
Contributors
The following have contributed to this page
