Prognostic model for pancreatic cancer based on machine learning of routine slides and transcriptomic tumor analysis
Menée à partir d'échantillons d'adénocarcinomes canalaires du pancréas prélevés sur 591 puis 302 patients, cette étude évalue la performance d'un algorithme d'apprentissage automatique, utilisant des données d'images de lames histologiques et des données transcriptomiques, pour établir un pronostic
Background : Prognostication for pancreatic ductal adenocarcinoma (PDAC) using histologic images is difficult due to tumor heterogeneity. We developed an artificial intelligence (AI) model to predict postoperative recurrence using histologic image patches.
Methods : We included 591 patients with resected PDAC to train an AI model for recurrence prediction at 12 or 24 months and validated it using external cohorts (n = 302 in total). Image patches from hematoxylin and eosin-stained slides were clustered via uniform manifold approximation and projection (UMAP) and used to train a random forest model. Predictive performance was evaluated using area under the receiver operating characteristic curve (AUC). Gene expression analysis was conducted to characterise survival-related clusters.
Results : Seventeen patch clusters were identified. Two were linked to high recurrence risk, and one to low risk. In external validation, the model achieved an AUC of up to 0.792. The random forest score independently predicted recurrence. Greater heterogeneity in patch composition correlated with shorter time to recurrence (P < 0.01). High-risk clusters showed elevated CSF3R expression; the low-risk cluster showed increased IGFBP3 expression.
Conclusions : Our AI model, using only archival histologic slides, accurately predicted postoperative recurrence in PDAC and revealed image features linked to outcomes and gene expression.
British Journal of Cancer , article en libre accès, 2026