3D CT foundation models fail generalization test in cancer recurrence study

A new arXiv paper (accepted at AIiH 2026) from researchers Bilel Guetarni, Feryal Windal, David Pasquier, and Halim Benhabiles systematically benchmarks 3D CT foundation models for predicting recurrence-free survival in head and neck cancer. The study addresses a critical gap: as these models become widely available, it's unclear whether their learned representations generalize across diverse clinical settings or require task-specific adaptation. Using two public datasets totaling 3,644 patients, the authors compared several 3D CT foundation models against traditional radiomics, which suffers from reproducibility issues and sensitivity to acquisition protocol variations. They also evaluated unsupervised adaptation strategies and different modality fusion mechanisms to see if these could unlock better performance.

Key findings reveal a persistent difficulty in identifying imaging features that generalize consistently across different distributions. Significant performance drops on external validation cohorts indicate that current 3D CT foundation models struggle with domain shift, even after adaptation. The most accurate approach involved fusing imaging features with clinical data, outperforming imaging-only models. However, achieving universal generalization across varied clinical contexts remains unsolved. For medical AI practitioners, this underlines that foundation models are not drop-in replacements for established radiological workflows—careful validation, adaptation, and multimodal integration are still required before deployment in real-world prognostic prediction.