Nasopharyngeal carcinoma (NPC) is a malignancy with a high cure rate when identified early, yet its detection remains problematic due to the complex anatomy of the nasopharynx and the resemblance of malignant features to benign conditions. These diagnostic challenges are particularly significant in primary care environments with limited resources and less experienced clinicians. A recent national, multicentre study in China addressed this gap by developing and validating an artificial intelligence system—the Swin Transformer-based Nasopharyngeal Diagnostic (STND)—to assist otolaryngologists in accurately identifying NPC, benign hyperplasia and normal nasopharynx using endoscopic images.
Developing a Multicentre Diagnostic Model
The development phase of the STND system involved collecting 27,362 endoscopic images from 15,521 patients across eight specialised NPC centres. These included biopsy-confirmed cases of NPC, benign hyperplasia and normal tissues. The model was designed using the Swin Transformer architecture, chosen for its capacity to analyse high-resolution visual data effectively. A classification output head was included to support accurate diagnostic categorisation.
The images were split into training, validation and internal testing datasets, with proportions of approximately 70%, 20% and 10%, respectively. The internal validation dataset demonstrated high performance with an AUC of 0.99 for both normal versus abnormal and malignant versus non-malignant classifications. Sensitivity and specificity were also strong, exceeding 93% in both discrimination tasks. These results provided confidence in the model’s robustness and readiness for testing in less controlled environments.
To evaluate generalisability, a prospective external validation was conducted using 1,885 images from ten additional hospitals situated in high-incidence regions for NPC. The system retained strong diagnostic performance, achieving an AUC of 0.95, a sensitivity of 91.6% and a specificity of 86.1%. These figures demonstrated that the model could maintain high accuracy despite the heterogeneity of data from different institutions, imaging equipment and clinical practices.
Clinical Evaluation through Reader Study
The model’s real-world utility was assessed through a multireader, multicase study involving four expert otolaryngologists and 24 primary care otolaryngologists. Each participant reviewed 400 anonymised endoscopic images twice—first unaided and then with STND support—separated by a two-week washout period to mitigate recall bias. Images were evenly distributed across NPC, benign hyperplasi and normal categories, and were drawn from the external validation set to represent realistic diagnostic scenarios.
The findings highlighted a clear improvement in diagnostic performance, particularly among primary care otolaryngologists. Their accuracy improved by 7.9%, from 83.4% to 91.2%, with notable gains in sensitivity, specificity and AUC for both abnormality and malignancy detection. Expert otolaryngologists also showed improvements, especially in classifying malignancies, with accuracy increasing by 6.2% and AUC by 0.06.
Must Read: Enhancing MRI Imaging in Nasopharyngeal Carcinoma Using ARMS DWI
Subgroup analyses provided further insights. Clinicians with less than 10 years’ experience saw the most pronounced benefits, with specificity in abnormality detection improving by 17.7%. The effect of hospital tier and endoscopy type was also explored, revealing that experience level, more than institutional resources, influenced diagnostic performance. Additionally, the use of STND significantly reduced image interpretation time from a mean of 6.7 seconds to 5.0 seconds per image, supporting its integration into fast-paced clinical workflows.
Practical Implications and Considerations
STND is particularly effective in reducing unnecessary biopsies by enhancing diagnostic specificity, a common issue in primary hospitals where cautious over-diagnosis often occurs to avoid missing cancers. The AI assistance helped less experienced clinicians reach specificity levels comparable to those of experts, reducing the psychological and physical burden on patients.
To address the ethical concerns surrounding AI integration, the developers implemented strong data privacy protections, including encryption and de-identification. They also advocated for offline operation to avoid cybersecurity threats and proposed institutional ethics committees to oversee implementation. The model’s training data were curated with attention to diversity in demographics, locations and imaging equipment, aiming to minimise bias and ensure equitable diagnostic outcomes.
Nevertheless, limitations remain. Performance varied across centres due to differences in imaging quality and protocols. The model relies on visible features in static images, limiting its utility for detecting submucosal cancers or those with subtle manifestations. Future research will focus on integrating more advanced imaging techniques and video data to improve sensitivity, especially for early-stage lesions that are not visible through conventional endoscopy.
The STND system demonstrates strong potential to enhance NPC diagnosis through AI-assisted endoscopy, particularly in primary care settings. By improving accuracy, reducing reliance on invasive procedures and shortening diagnostic time, the system contributes meaningfully to earlier intervention and better patient outcomes. With further development, including video analysis and integration with advanced imaging technologies, STND could play a pivotal role in strengthening NPC screening and diagnosis in high-incidence regions, supporting national and international cancer care goals.
Source: The Lancet Digital Health
Image Credit: iStock
