Throat AI - An Intelligent System For Detecting Foreign Objects In Lateral Neck X-Ray Images

• Published in: International Conference on Artificial Intelligence and Big Data (Online) pp. 717 - 722
• Main Authors: Baddewithana, Pasan, Krishara, Jenny, Yapa, Kanishka
• Format: Conference Proceeding
• Language: English
• Published: IEEE 23.05.2025
• Subjects: Accuracy; artificial intelligence; Biomedical imaging; Computational modeling; Deep learning; image processing; lateral neck x-ray; Neck; Testing; Training; X-rays; YOLO
• ISSN: 2769-3554
• DOI: 10.1109/ICAIBD64986.2025.11082052

Foreign Object ingestion is a commonly encountered medical condition within the Ear, Nose, and Throat clinical domain. Timely and accurate detection of such objects is vital, as it often guides the need for surgical intervention. Among the available imaging techniques, lateral neck X-rays are the most widely used radiographs to visualize and assess the presence of FOs in the throat. However, manual interpretation of these images can be time-consuming and subject to human error, potentially leading to misdiagnosis or delayed treatment. This research presents a deep learning-based software solution, deployable via web and mobile platforms, aimed at assisting medical professionals with the automated detection of FOs in lateral neck X-rays. The system leverages state-of-the-art YOLO object detection models, specifically evaluating novel versions such as YOLO-NAS-s, YOLOv11s, and YOLOv8s-OBB to ensure high detection accuracy and deployment efficiency. The best-performing model, YOLO-NAS-s, achieved a validation accuracy of 96.3%. For deployment, the model was hosted on the Roboflow platform and accessed via a FastAPI-based middleware server. Performance evaluation showed an average inference time of approximately 2 seconds and a memory footprint of around 100 MB on standard computing hardware, demonstrating its suitability for integration into resource-constrained clinical environments. This setup highlights the system's lightweight design and real-world applicability. Training, evaluation, and testing of the deep learning models were conducted using a dataset curated from public local healthcare institutions and online medical imaging repositories.