[Development and preliminary validation of an integrated artificial intelligence system for quality-control...

Objective: To develop ENDOVISTA-ENT, an integrated artificial intelligence system for quality-control assistance and lesion recognition during nasopharyngolaryngoscopy, and to evaluate its model performance and clinical utility. Methods: This retrospective study included 2 365 eligible patients from two centers. Of these, 1 562 patients from the First Affiliated Hospital of Sun Yat-sen University were used for model development and internal testing, and 803 patients from Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, were used for external validation. Model 1 identified whether endoscopic images were acquired inside or outside the body. Model 2 identified 11 standard anatomical sites. Model 3 localized lesions and classified them as benign or malignant. In addition, 200 pathologically confirmed cases were randomly selected from the test set. Six physicians with different levels of experience performed independent and AI-assisted interpretation. The real-time performance of the system was also assessed. Results: The internal and external accuracies of Model 1 were 99.44% and 99.84%, respectively. The corresponding accuracies of Model 2 were 96.09% and 94.73%, and the areas under the receiver operating characteristic curve (AUCs) for individual anatomical sites ranged from 0.99 to 1.00. For Model 3, the internal and external precision values were 84.11% and 82.73%, and the recall values were 80.60% and 78.31%, respectively. The AUCs for malignant lesions were 0.986 and 0.968. For early-stage nasopharyngeal, laryngeal, and hypopharyngeal cancers, the AUCs ranged from 0.878 to 0.932 across the internal test and external validation sets. AI assistance increased physicians' overall accuracy from 78.50% to 88.20% (χ²=40.37, P<0.001). The single-frame detection time of Model 3 was (15.2±3.4) ms. The system processed (28.5±2.1) frames/s, with an end-to-end latency of <40 ms. Conclusion: ENDOVISTA-ENT demonstrated good model performance and real-time processing capability, improved physicians' lesion interpretation, and provided a basis for further clinical validation.