Preoperative prediction of perineural invasion in intrahepatic cholangiocarcinoma with interpretable machine learning...

To develop radiomics and deep learning (DL) based interpretable models using MRI for preoperative prediction of perineural invasion (PNI) in intrahepatic cholangiocarcinoma (ICC).

A total of 165 pathologically confirmed ICC patients with preoperative MRI were retrospectively enrolled from two centers (center1, training set, n = 115; validation set, n = 14; internal test set, n = 15; center 2, external test set, n = 21). Radiomics and DL models were constructed for single-phase (pre-contrast, arterial phase, portal venous phase, hepatobiliary phase [HBP]) and multi-phase MRI using the Shukun AI platform and PNI-MambaNet. Model performance was evaluated with the area under the receiver operating characteristic curve (AUC). Gradient-weighted class activation mapping (Grad-CAM) heatmaps visualized the regions prioritized by the DL models.

The PNI positive rate was 42.4 % (61/144) and 28.6 % (6/21) in the two centers. Radiomics HBP models achieved the highest AUC in the internal test set, while multi-phase model performed best in the external test set (AUC: HBP, 0.778 and 0.733 for the internal and external test sets, respectively; multi-phase, 0.759 and 0.778). For DL models, multi-phase model achieved the highest AUC in the internal test set, while HBP model performed best in the external test set (AUC: HBP, 0.926 and 0.856; multi-phase, 0.944 and 0.844). DL models outperformed radiomics models in the external test set, with Grad-CAM visualizing tumor margin regions as the interest area.

DL models based on MRI effectively predict PNI in ICC, with visualizations enhancing clinical interpretability and potential application.