Fine-tuned Vision Transformers and YOLOv11 for precise detection of pediatric Adenoid Hypertrophy

This study introduces an advanced AI-driven framework for the automated detection of pediatric Adenoid Hypertrophy (AH) in lateral nasopharyngeal radiographs, utilizing a hybrid architecture ViT-CNN model that integrates Vision Transformers (ViT) and Convolutional Neural Networks (CNN). Additionally, YOLOv11 was employed for precise segmentation of adenoid structures. The framework incorporates fine-tuning techniques and evaluates performance under conditions with and without data augmentation to ensure a comprehensive analysis of their capabilities. The study utilized a dataset of 900 lateral nasopharyngeal radiographs from pediatric patients, representing diverse demographic and clinical characteristics. The models achieved exceptional diagnostic accuracy, with 100% precision and high Receiver Operating Characteristic Area Under the Curve (ROC-AUC) values, indicating a robust ability to distinguish between diagnostic categories. This level of accuracy suggests significant potential for reducing diagnostic errors, improving diagnostic turnaround times, and enhancing efficiency in clinical workflows, particularly in pediatric care. Unlike existing methods, which rely heavily on manual landmark identification and exhibit poor generalization across varied datasets, this framework ensures precise segmentation and robust classification, overcoming these limitations. The framework is clinically relevant as it streamlines radiological workflows, minimizes the workload for radiologists, and provides reliable automated detection of AH in children. While achieving impressive results, potential challenges such as dataset imbalances and computational demands were identified. Future efforts will focus on synthetic data generation and real-time optimization to enhance the framework’s clinical applicability.