Biologically interpretable deep learning–derived MRI phenotypes reveal lymph node involvement and neoadjuvant therapy response in intrahepatic cholangiocarcinoma

BACKGROUND AND AIMS: Accurate N-staging of intrahepatic cholangiocarcinoma (iCCA) remains challenging using noninvasive approaches. We aimed to develop a model to refine lymph node (LN) involvement stratification and inform therapeutic consideration. APPROACH AND RESULTS: This study enrolled a discovery cohort (n=682), an internal test cohort from the FU-iCCA (n=204), and an external multicenter cohort (n=88) for model development, and a neoadjuvant therapy (NAT) cohort (n=145) for therapeutic evaluation of the model. A SwinU-CliRad framework was constructed by integrating Swin UNEt TRansformers (Swin UNETR)-based magnetic resonance imaging-derived outputs of LN involvement with clinicoradiological features. Correlations between SwinU outputs and tumor multi-omics profiles were explored. The SwinU-CliRad model achieved areas under the curve (AUCs) of 0.932, 0.867, and 0.888 in LN risk stratification, and outperformed radiologist-based assessments by correcting more misclassifications than it introduced across the discovery, internal, and external test cohorts (18.8% vs. 7.3%, 18.1% vs. 4.9%, and 17.0% vs. 5.7%), respectively. In the NAT cohort, patients classified as high LN-involved risk by the SwinU-CliRad exhibited lower residual viable tumor rates than those with low LN-involved risk, with higher rates of pathological complete response (12.0% vs. 4.2%) and major pathological response (14.0% vs. 8.4%). SwinU outputs were associated with KRAS mutations, MUC5AC overexpression, and the large-duct histological subtype. Single-cell RNA sequencing analysis linked LN involvement to an immune-suppressive stroma tumor microenvironment. CONCLUSIONS: The SwinU-CliRad model can serve as a biologically interpretable tool for LN risk stratification in iCCA surgical candidates, with high-risk patients identified by the model potentially deriving benefit from NAT.