Optimized deep learning model for diagnosing tonsil and adenoid hypertrophy through X-rays.

Objective: To explore the application of a deep learning model based on lateral nasopharyngeal X-rays in diagnosing tonsillar and adenoid hypertrophy. Methods: A retrospective study was conducted using DICOM images of lateral nasopharyngeal X-rays from pediatric outpatients aged 2-12 at our hospital from July 2014 to July 2024. The study included patients exhibiting varying degrees of respiratory obstruction symptoms (disease group). Initially, 1006 images were collected, but after excluding low-quality images and standardizing the imaging phase, 819 images remained. These images were divided into training and validation sets in an 8:2 ratio. The independent test set is consisted of 484 images. We delineated the target areas for tonsils and adenoids and used a YOLOv8n-based model for object detection and use various convolutional neural network models to classify the cropped images, assessing the severity of tonsillar and adenoid hypertrophy. We compared the performance of these models on the training and validation sets using metrics such as ROC-AUC, accuracy, precision, recall, and F1 score. Results: The combined model, incorporating YOLOv8 for object detection and secondary classification, demonstrated excellent performance in diagnosing tonsillar and adenoid hypertrophy, significantly improving diagnostic accuracy and consistency. The ResNet18 model, due to its lightweight nature and minimal computational resource requirements, performed exceptionally well in the YOLOv8-ResNet fusion model for detecting and classifying tonsils and adenoids, making it our preferred model. Conclusion: The deep learning model combining YOLOv8n and ResNet18 based on lateral nasopharyngeal X-rays demonstrates significant advantages in diagnosing pediatric tonsillar and adenoid hypertrophy.