Robust multicentre detection and classification of colorectal liver metastases on CT: application of foundation models.

Colorectal liver metastases (CRLM) remain a major cause of cancer-related mortality. Early, reliable detection on CT imaging is essential for curative treatment planning, yet the diagnostic performance of AI models often declines across scanners and institutions, limiting clinical generalizability. In this study, we developed and evaluated a multi-center foundation-model-based AI pipeline for patient-level classification and lesion-level detection of CRLM on contrast-enhanced CT. Using data from the EuCanImage consortium (n = 2437) and TCIA_CRLM (n = 197, all CRLM), we benchmarked several pretrained foundation models and identified UMedPT as the optimal encoder. The final model achieved an AUC of 0.89 and a sensitivity of 0.82 on the EuCanImage test set, with a sensitivity of 0.85 on the external TCIA cohort. Excluding the most uncertain 20% of cases improved AUC to 0.90 and balanced accuracy to 0.85. Decision-curve analysis indicated superior net benefit over "treat-all" and "treat-none" strategies for threshold probabilities between 0.35 and 0.75. The lesion-level detector identified 69.1% of lesions overall, increasing substantially with lesion size. Grad-CAM maps showed strong correspondence between attention regions and metastases in high-confidence predictions. These findings demonstrate that foundation-model-based pipelines enable robust, generalizable, and interpretable solutions for CRLM detection and classification in multi-center CT imaging.