Real-Time-Capable Detection of Glottic, Supraglottic and Hypopharyngeal Lesions Using Artificial Intelligence During Flexible Endoscopy

Background/Objectives: Supraglottic and hypopharyngeal carcinomas are aggressive malignancies that are often diagnosed at advanced stages. Timely recognition of these malignancies is influenced by many factors, such as endoscope quality and experience. This study evaluated the potential of artificial intelligence (AI) to support real-time detection and classification of such lesions during flexible endoscopy in the outpatient clinic. Methods: A previously developed deep learning (DL) algorithm was extended from a glottic lesion model to the unified localization and classification of glottic, supraglottic, and hypopharyngeal lesions during flexible endoscopy. Lesion frames were extracted from endoscopy videos obtained at two head and neck oncology centers and one secondary referral center between 2012 and 2024. These frames were annotated and labeled based on histopathological or clinically confirmed reference diagnoses. The primary outcome was the unified model’s performance in detection of lesions per frame. After training (70% of data), the positive predictive value (precision) and sensitivity (recall) of this model were calculated on an independent test set (30% of data), stratified by subsite and tumor (T-) classification. Secondly, the model’s binary classification performance (benign or malignant) was evaluated. Results: From 490 supraglottic and hypopharyngeal endoscopy videos, 40,059 frames with a benign or malignant lesion were extracted and added to the 56,036 glottic lesion frames in the database, comprising 1336 lesion videos and 123 healthy control videos (total n = 1459). On the test set, the model achieved a detection precision of 92.1% (95% CI: 90.9–93.2) and a recall of 73.3% (95% CI: 69.8–76.6). Detection performance increased with higher T-stage. Among correctly detected lesion instances in a malignancy-dominated test set (65.3% of lesion instances), sensitivity for malignancy detection was 94.3% (95% CI: 91.8–96.5). Combined end-to-end performance for correct malignant lesion detection and classification was estimated at 69.1% (object-level recall 73.3% × classification sensitivity 94.3%). Evaluation of 33 lesion-free videos demonstrated a mean frame-level specificity of 46.6% and a median of 19 false positive detections per video. Conclusions: This is the first study to report a DL model for real-time endoscopic detection and classification of benign and malignant laryngeal and pharyngeal lesions. The developed model showed promising lesion detection and cancer classification performance in the evaluated test set. T1 tumor detection remains an important limitation, particularly because early-stage detection is a primary aim of AI-assisted endoscopy. Further model testing is required in real-world lesion prevalence settings, where external validation and clinical usability should be investigated.