(503) Deep Learning-Assisted Endoscopic Diagnosis of Laryngopharyngeal Reflux by Recognizing Signs of Chronic Inflammation

Introduction: Laryngopharyngeal reflux disease (LPR) is a common condition, but its current diagnostic methods are costly and invasive, creating a need for a more efficient alternative based on endoscopic signs of chronic inflammation (LPR-CI). However, accurately identifying these signs is challenging due to complex tissue variations and subjective visual assessment, prompting the question of whether an automated, objective detection method can be developed.

Methods: Thus, we proposed a deep-learning approach, i.e., the balanced region self-attention reconstruction-based LPR diagnosis (BRSAR-LPRD) model, for LPR-CI sign detection. This BRSAR-LPRD model employs adaptive pixel class-balanced loss to manage pixel-level label imbalance and reduce overfitting from small-region segmentation and limited annotations. Cross-validation employs both internal and external datasets, incorporating a comparison with otolaryngology diagnoses

Results: The validation results demonstrated that BRSAR-LPRD achieved an accuracy of 0.91 and an area under the curve of 0.88, surpassing that of general clinicians and matching that of chief otolaryngologists (p < 0.05). The validity of the model was further supported by pathophysiological evidence, including higher Grade 2 CD3 expression (p=0.037).

Conclusions: Consequently, this study introduces the first expert-level methodology for identifying suspected LPR, thereby furnishing physicians with a simple, rapid, and cost-effective screening or assistive diagnostic tool.