3D CT Foundation Models Show Limited Generalization in Head and Neck Cancer Recurrence Prediction
A recent study available on arXiv (2608.00071) assesses multiple 3D CT foundation models aimed at predicting recurrence-free survival in patients with head and neck cancer, utilizing two public datasets comprising 3,644 individuals. This research examines different adaptation strategies and methods for modality fusion, highlighting ongoing challenges in consistently identifying features across varying imaging distributions. The results indicate that although 3D CT foundation models present a promising alternative to traditional radiomics—known for its reproducibility challenges and sensitivity to variations in acquisition protocols—their representations may not effectively generalize in diverse clinical environments without specific adaptations for each task. The research emphasizes the importance of thorough evaluation and customization of these models for subsequent applications.
Key facts
- Study benchmarks 3D CT foundation models for head and neck cancer recurrence prediction.
- Uses two public datasets totaling 3,644 patients.
- Evaluates various adaptation strategies and modality fusion mechanisms.
- Findings show difficulty in identifying features that generalize across imaging distributions.
- 3D CT foundation models are compared to traditional radiomics.
- Traditional radiomics suffers from reproducibility issues and sensitivity to acquisition protocol variations.
- The study is published on arXiv with identifier 2608.00071.
- The research addresses the need to evaluate generalization of foundation models in clinical settings.
Entities
Institutions
- arXiv