Artificial intelligence-assisted detection of nasopharyngeal carcinoma on endoscopic images: a national, multicentre, model development and validation study.

Abstract

Background: Nasopharyngeal carcinoma is highly curable when diagnosed early. However, the nasopharynx's obscure anatomical position and the similarity of local imaging manifestations with those of other nasopharyngeal diseases often lead to diagnostic challenges, resulting in delayed or missed diagnoses. Our aim was to develop a deep learning algorithm to enhance an otolaryngologist's diagnostic capabilities by differentiating between nasopharyngeal carcinoma, benign hyperplasia, and normal nasopharynx during endoscopic examination.

Methods: In this national, multicentre, model development and validation study, we developed a Swin Transformer-based Nasopharyngeal Diagnostic (STND) system to identify nasopharyngeal carcinoma, benign hyperplasia, and normal nasopharynx. STND was developed with 27 362 nasopharyngeal endoscopic images (10 693 biopsy-proven nasopharyngeal carcinoma, 7073 biopsy-proven benign hyperplasia, and 9596 normal nasopharynx) sourced from eight prominent nasopharyngeal carcinoma centres (stage 1), and externally validated with 1885 prospectively acquired images from ten comprehensive hospitals with a high incidence of nasopharyngeal carcinoma (stage 2). Furthermore, we did a fully crossed, multireader, multicase study involving four expert otolaryngologists from four regional leading nasopharyngeal carcinoma centres, and 24 general otolaryngologists from 24 geographically diverse primary hospitals. This study included 400 images to evaluate the diagnostic capabilities of the experts and general otolaryngologists both with and without the aid of the STND system in a real-world environment.

Findings: Endoscopic images used in the internal study (Jan 1, 2017, to Jan 31, 2023) were from 15 521 individuals (9033 [58·2%] men and 6488 [41·8%] women; mean age 47·6 years [IQR 38·4-56·8]). Images from 945 participants (538 [56·9%] men and 407 [43·1%] women; mean age 45·2 years [IQR 35·2- 55·2]) were used in the external validation. STND in the internal dataset discriminated normal nasopharynx images from abnormalities (benign hyperplasia and nasopharyngeal carcinoma) with an area under the curve (AUC) of 0·99 (95% CI 0·99-0·99) and malignant images (ie, nasopharyngeal carcinoma) from non-malignant images (ie, benign hyperplasia and normal nasopharynx) with an AUC of 0·99 (95% CI 0·98-0·99). In the external validation, the system had an AUC for the detection of nasopharyngeal carcinoma of 0·95 (95% CI 0·94-0·96), a sensitivity of 91·6% (95% CI 89·3-93·5), and a specificity of 86·1% (95% CI 84·1-87·9). In the multireader, multicase study, the artificial intelligence (AI)-assisted strategy enhanced otolaryngologists' diagnostic accuracy by 7·9%, increasing from 83·4% (95% CI 80·1-86·7, without AI assistance) to 91·2% (95% CI 88·6-93·9, with AI assistance; p<0·0001) for primary care otolaryngologists. Reading time per image decreased with the aid of the AI model (mean 5·0 s [SD 2·5] vs 6·7 s [6·0] without the model; p=0·034).

Interpretation: Our deep learning system has shown significant clinical potential for the practical application of nasopharyngeal carcinoma diagnosis through endoscopic images in real-world settings. The system offers substantial benefits for adoption in primary hospitals, aiming to enhance specificity, avoid additional biopsies, and reduce missed diagnoses.

Funding: New Technologies of Endoscopic Surgery in Skull Base Tumor: CAMS Innovation Fund for Medical Science; Shanghai Science and Technology Committee Foundation; Shanghai Municipal Key Clinical Specialty; National Key Clinical Specialty Program; and the Young Elite Scientists Sponsorship Program.