Artificial intelligence-assisted real-time nasopharyngeal cancer diagnostic model enhances rhinologist performance: a prospective multi-reader study

Le résumé fourni par la source

Background Nasopharyngeal carcinoma (NPC) poses significant diagnostic challenges due to the anatomical complexity of the nasopharynx and reliance on endoscopic visual interpretation, often leading to delayed detection and unnecessary biopsies. Although artificial intelligence (AI) algorithms have shown promise in enhancing endoscopic cancer diagnosis, their real-world impact on clinician diagnostic performance remains insufficiently characterized.Methods In this prospective, multi-reader study, 47 clinicians, including experts, residents, and trainees, interpreted 200 nasoendoscopic images from 100 patients with histopathologically confirmed NPC or benign lesions. Each participant completed two diagnostic sessions: an unassisted evaluation and an AI-assisted assessment using a real-time lesion-annotating model (NPC-SDNet), with a 4-week washout period between sessions.Results Without AI support, the overall diagnostic accuracy was 73.6% (95% CI: 70.1–77.0%), with a sensitivity of 76.1% (95% CI: 70.4–81.4%) and a specificity of 69.9% (95% CI: 63.8–76.0%). AI assistance significantly improved accuracy to 85.6% (95% CI: 83.1–87.6%, p < 0.001), sensitivity to 90.1% (95% CI: 86.6– 92.9%, p < 0.001), and specificity to 79.1% (95% CI: 75.6–82.7%, p < 0.001). Subgroup analysis revealed the greatest improvements among trainees (64.8% vs 83.5%, p < 0.001) and residents (77.2% vs 84.9%, p = 0.003). Moreover, AI integration substantially reduced median image interpretation time from 1411.7 to 818.5 s (p < 0.001).Conclusions AI-assisted nasoendoscopic evaluation significantly enhances diagnostic accuracy, efficiency and interobserver consistency, particularly among less-experienced clinicians. These findings support the clinical integration of real-time AI tools to augment NPC recognition and standardize diagnostic performance across varying expertise levels.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.