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

Area of Science:

• Artificial Intelligence in Medicine
• Oncology
• Medical Imaging

Background:

• Nasopharyngeal carcinoma (NPC) diagnosis relies heavily on endoscopic examination and biopsy.
• Accurate and early detection of NPC is crucial for patient outcomes.
• Existing diagnostic tools can be limited by inter-observer variability and accessibility.

Purpose of the Study:

• To develop and validate an interpretable endoscopic intelligence system for NPC diagnosis.
• To provide clinical decision support through an integrated "screening-segmentation-biopsy localization" model.
• To enhance diagnostic accuracy from lesion detection to precise tissue sampling.

Main Methods:

• A multi-center study collected 12,323 nasopharyngeal endoscopic images (2018-2022).
• An encoder-multi-head decoder architecture was used with a multi-task joint loss function.
• Model performance was validated using single-center training and multi-center testing, with permutation and DeLong tests for statistical significance.

Main Results:

• Internal testing showed high performance: AUC 0.998, sensitivity 91.56%, specificity 99.61%.
• External validation demonstrated generalizability: AUC 0.957, sensitivity 97.63%, specificity 76.79%.
• AI sensitivity (92.00%) outperformed most physicians (78.40%), while senior physicians had higher specificity (95.00%). Segmentation Dice coefficient was 0.86, with 67.7% biopsy site accuracy.

Conclusions:

• A highly sensitive, multi-task AI system for NPC diagnosis was successfully developed.
• The AI system effectively assists in lesion detection, segmentation, and biopsy site localization.
• This AI tool complements physicians' diagnostic capabilities, potentially improving NPC management.