A confidence-based, artificial intelligence pathology model for diagnosis of intrahepatic cholangiocarcinoma
Menée à partir de données de 5 centres européens spécialisés portant sur 544 patients atteints d'un cholangiocarcinome intrahépatique ou de métastases issues de cancers extrahépatiques puis validée auprès de 161 patients supplémentaires traités en France, en Inde ou en Corée, cette étude compare l'efficacité de trois architectures d'apprentissage automatique pour établir un diagnostic puis évalue la performance du modèle AI2CCA pour distinguer une métastase hépatique d'un cholangiocarcinome intrahépatique à partir de lames histologiques numérisées
Background : Intrahepatic cholangiocarcinoma (ICCA) is a rare but highly lethal adenocarcinoma arising within the hepatic parenchyma. Diagnosis presents a significant clinical challenge as the histological features of ICCA substantially overlap with those of metastatic liver cancers. This diagnostic ambiguity often necessitates extensive and costly exclusionary investigations, such as upper and lower gastrointestinal endoscopy, to rule out an occult primary site. Consequently, this process results in treatment delays and an increased financial burden on healthcare systems.
Patients and Methods : We retrospectively analyzed 544 patients across five European centers, comprising cases of either ICCA or metastases from extrahepatic cancers. Three deep-learning architectures utilizing foundation models were investigated: Ctranspath/HistoBistro, UNI/CLAM, and CONCH/TITAN. Performance was assessed using the Area Under the Receiver Operating Characteristic curve (AUROC) and the False Positive Rate (FPR). Furthermore, we implemented a confidence estimation system using the Generalized-ODIN (G-ODIN) approach, utilizing predictive entropy as a metric. The final model, designated AI2CCA, was prospectively validated in 161 patients across four international centers in France, India, and Korea.
Results : In the retrospective test set, the CONCH/TITAN architecture yielded the best performance (AUROC: 0.840). Predictive entropy derived via G-ODIN was significantly higher in misclassified cases, validating its utility as a confidence metric. Implementation of confidence thresholding improved the AUROC to 0.958 with an FPR of 0, while retaining 46% of samples for high-confidence prediction. In prospective validation, AI2CCA achieved AUROCs of 1.00 and 0.965 in the French and Asian cohorts, respectively (with only one misclassified case in the Asian series).
Conclusion : Collectively, our study demonstrates the real-world clinical utility of a confidence-based AI biomarker for assisting in the diagnosis of liver cancer. By accurately discriminating ICCA from metastasis, this tool offers the potential to reduce unnecessary investigations and accelerate therapeutic decision-making.
Annals of Oncology , article en libre accès, 2026