[Integrated multitask visual-guided system for nasopharyngeal carcinoma detection using nasopharyngeal endoscopy: a...

Objectives: To develop and validate an interpretable endoscopic intelligence system for assisting in the diagnosis of nasopharyngeal carcinoma (NPC). Through an integrated model of "screening-segmentation-biopsy localization", to provide clinical decision support from initial lesion detection to precise tissue sampling. Methods: A multi-center model development and validation study was conducted. A total of 12 323 nasopharyngeal endoscopic images from three medical institutions (including Eye & ENT Hospital of Fudan University, the First Affiliated Hospital of Guangxi Medical University and Guangxi Zhuang Autonomous Region People's Hospital) from 2018 to 2022 were collected. Model performance was validated using a single-center training and multi-center testing approach. An encoder-multi-head decoder architecture was employed, optimized via a multi-task joint loss function to simultaneously perform classification, segmentation, and biopsy site localization. Permutation testing (10 000 permutations) was used to evalute the significance of sensitivity and specificity, and the DeLong test was applied for receiver operating characteristic (ROC) curve comparison. Results: The model exgibited strong screening performance in internal testing: area under curve (AUC) was 0.998 (95%CI: 0.998-0.999), sensitivity was 91.56% (95%CI: 90.64%-92.40%), and specificity was 99.61% (95%CI: 99.37%-99.80%). External validation confirmed strong generalizability with an AUC of 0.957 (95%CI: 0.949-0.964), sensitivity of 97.63% and specificity of 76.79%. In human-AI comparison, the model's sensitivity (92.00%, 95%CI: 86.79%-96.00%) significantly outperformed most physicians, e.g., higher than senior physicians' average sensitivity (model 92.00% vs 78.40%, P<0.05), while senior physicians exhibited superior specificity (95.00% vs model 90.00%). The segmentation task yielded a Dice coefficient of 0.86 (95%CI: 0.80-0.93), and the recommended biopsy sites were accurate in 67.7% of cases. Conclusion: This study successfully develops a highly sensitive multi-task AI system for NPC diagnosis, which complements physicians' diagnostic strengths.