MRI Deep Learning Model Distinguishes Liver Cancer Types Before Surgery

The team reports that their most comprehensive model, integrating multi-phase MRI features with clinical predictors, achieved an area under the curve (AUC) of 0.887 (95% CI, 0.819–0.954) in an internal validation cohort and 0.908 (95% CI, 0.811–1.000) in an independent external validation cohort.

The internal development cohort included 334 surgically confirmed patients, 231 with HCC and 103 with MF-ICC, divided at the patient level into a training set of 233 and an internal validation set of 101. An external validation cohort of 50 patients, 40 with HCC and 10 with MF-ICC, tested generalization. The researchers first constructed separate two-dimensional deep learning models based on arterial phase, portal venous phase, and delayed phase images, then developed fusion models combining them. They also built a multi-channel model by directly stacking the three-phase images along the channel dimension, and a clinical model from independent predictors identified through multivariate analysis, which was ultimately integrated with the best-performing deep learning model.

The multi-channel approach, feeding all three contrast phases simultaneously into the network, discriminated better than every single-phase and corresponding fusion model, reaching an AUC of 0.814 in internal validation. The full combination, a 2D plus multi-channel plus clinical model, performed best overall. Notably, however, the authors report that after correction for multiple comparisons, no statistically significant differences in AUCs remained among the tested models, a caution that tempers any claim that one architecture decisively outperforms the others.

The authors conclude that the fusion model integrating clinical predictors with multi-phase DCE-MRI deep learning features may provide a valuable preoperative decision-support instrument, while the modest external cohort, particularly only ten MF-ICC cases, and the absence of significant between-model differences after multiplicity correction mean the framework requires broader validation before routine clinical use.

Subject of Research: Deep learning on dynamic contrast-enhanced MRI combined with clinical modeling to distinguish hepatocellular carcinoma from mass-forming intrahepatic cholangiocarcinoma before surgery

Article Title: Dynamic contrast-enhanced MRI deep learning model combined with clinical model for noninvasive differentiation of hepatocellular carcinoma from mass-forming intrahepatic cholangiocarcinoma

Article References: Zhang, H., Guo, L., Jia, F., Liu, W., Zhang, M., Li, F., Zhou, J., & Wu, B. (2026). Dynamic contrast-enhanced MRI deep learning model combined with clinical model for noninvasive differentiation of hepatocellular carcinoma from mass-forming intrahepatic cholangiocarcinoma. Frontiers in Oncology, 16, Article 1936989. https://doi.org/10.3389/fonc.2026.1936989

Image Credits: AI Generated

DOI: 10.3389/fonc.2026.1936989

Keywords: deep learning, DCE-MRI, hepatocellular carcinoma, intrahepatic cholangiocarcinoma, liver cancer, medical imaging, diagnostic model, radiomics, clinical prediction, preoperative diagnosis, Dynamic, contrast-enhanced