Artificial intelligence-based pathological model for pan-cancer lymph node metastasis detection: a multicentre diagnostic study with retrospective and prospective validation
Menée en Chine à partir de lames histologiques de ganglions lymphatiques provenant de 9 256 patients (âge médian : 60 ans), cette étude multicentrique évalue la performance d'un modèle basé sur l'intelligence artificielle pour détecter des métastases ganglionnaires issues de différents types de cancers
Background : Accurate detection of lymph node metastasis is crucial for precise tumour staging and treatment planning. Conventional pathological examination can overlook lymph node micrometastasis, resulting in underdiagnosis and suboptimal clinical outcomes. This study aimed to develop a pan-cancer artificial intelligence diagnostic model (PanCAM) for detecting lymph node metastasis across cancer types.
Methods : In this multicentre diagnostic study, patients who had undergone tumour resection and lymph node dissection from 17 hospitals in China were included. The entire dataset included nine common and 24 rare cancers. Histological slides of resected lymph nodes were collected and scanned to generate whole slide images (WSIs). PanCAM was developed by use of supervised learning and incremental learning strategies and trained and internally validated retrospectively on WSIs from nine common cancers at Sun Yat-sen Memorial Hospital of Sun Yat-sen University (Guangzhou, China). Its generalisability was retrospectively validated by use of WSIs from 15 external hospitals in China and the publicly available CAMELYON16 dataset from the Netherlands, representing both common and rare cancers. Prospective validation was done in a multicentre study (NCT06517979) across nine hospitals in China. The primary outcome was diagnostic sensitivity for detecting lymph node metastasis. In a secondary analysis, we compared performance between PanCAM and pathologists.
Findings : Between Jan 1, 2013, and Nov 30, 2024, 9256 patients from 17 hospitals in China were included in the study (4735 [51·2%] men; 4521 [48·8%] women; median age 60 years [IQR 50–69]; 3486 [37·7%] with lymph node metastasis). The dataset comprised 1303 patients in the training set, 558 in the internal validation set, 6006 in the external validation sets, and 1389 in the prospective validation sets, totalling 69 502 images and 153 985 lymph nodes. The CAMELYON16 dataset consisted of 399 images. In the retrospective validation, the diagnostic sensitivity of PanCAM for detecting lymph node metastasis ranged from 0·97 (95% CI 0·92–0·99) to 1·00 (0·98–1·00) across 16 hospitals, and was 0·96 (0·92–0·99) on the CAMELYON16 dataset. In the prospective validation, PanCAM exhibited sensitivity ranging from 0·93 (0·78–0·99) to 1·00 (0·98–1·00) across nine hospitals. Despite being trained solely on images from common cancers, PanCAM achieved a sensitivity of 0·98 (0·95–1·00) for rare cancers in both retrospective and prospective validations. At the patient level, PanCAM identified 120 additional patients with lymph node metastasis who were missed by pathologists in the retrospective validation and 21 additional cases in the prospective validation.
Interpretation : PanCAM provided a generalisable solution for detecting lymph node metastasis across cancer types. With high sensitivity and robust performance, the model could assist pathologists in diagnosing lymph node metastasis, improving diagnostic accuracy, supporting treatment decision making, and ultimately enhancing patient outcomes.
Funding : National Natural Science Foundation of China, National Science and Technology Major Project, Science and Technology Projects in Guangzhou, Guangdong Provincial Clinical Research Centre for Urological Diseases, and Science and Technology Planning Project of Guangdong Province.
The Lancet Digital Health , article en libre accès, 2026