Early detection of nasopharyngeal carcinoma greatly increases the chances of successful treatment. However, the nasopharynx’s concealed position and the visual overlap of its lesions with those of various benign disorders frequently complicate accurate identification, leading to delayed or missed diagnoses. This project aimed to build a deep learning tool to help otolaryngologists distinguish nasopharyngeal carcinoma, benign hyperplasia, and healthy nasopharyngeal tissue during routine endoscopic procedures.Within this nationwide, multi-institutional effort focused on model creation and testing, researchers designed the Swin Transformer-based Nasopharyngeal Diagnostic (STND) framework to categorize endoscopic views into nasopharyngeal carcinoma, benign hyperplasia, or normal tissue. Training relied on 27,362 endoscopic images (including 10,693 biopsy-verified nasopharyngeal carcinoma cases, 7,073 biopsy-verified benign hyperplasia cases, and 9,596 normal cases) collected across eight major nasopharyngeal carcinoma institutions (stage 1). Prospective external testing included 1,885 new images collected from 10 hospitals in regions with elevated disease rates (stage 2). A complete crossover multireader multicase experiment further included four specialists from top regional nasopharyngeal carcinoma facilities and 24 general otolaryngologists working in 24 varied primary care settings. This component analyzed 400 images to measure changes in diagnostic performance when clinicians used the STND system versus working independently in practical scenarios.Internal dataset images, obtained between January 1, 2017, and January 31, 2023, came from 15,521 people (9,033 men [58.2%] and 6,488 women [41.8%]; average age 47.6 years [IQR 38.4–56.8]). External validation used images from 945 individuals (538 men [56.9%] and 407 women [43.1%]; average age 45.2 years [IQR 35.2–55.2]). On internal data, STND reached an AUC of 0.99 (95% CI: 0.99–0.99) to distinguish normal nasopharynx from any abnormality (benign hyperplasia plus carcinoma) and an AUC of 0.99 (95% CI: 0.98–0.99) to identify malignant versus non-malignant cases. External validation yielded an AUC of 0.95 (95% CI: 0.94–0.96) for carcinoma detection, 91.6% sensitivity (95% CI: 89.3–93.5), and 86.1% specificity (95% CI: 84.1–87.9). In the reader experiment, AI support raised primary care otolaryngologists’ overall accuracy by 7.9%, moving from 83.4% (95% CI: 80.1–86.7) unaided to 91.2% (95% CI: 88.6–93.9) with assistance (P < 0.0001). Per-image interpretation time also fell from 6.7 s (SD: 6.0) to 5.0 s (SD: 2.5) when using the model (P = 0 . 034).The developed deep learning platform shows strong promise for everyday clinical deployment in recognizing nasopharyngeal carcinoma via endoscopic imaging. It delivers clear advantages for frontline hospitals by improving specificity, reducing unnecessary biopsies, and preventing missed malignancies.
