Comparing the Performance of Foundation Model Derived Embeddings with Traditional Approaches for Distant Metastasis Prediction in Head and Neck Cancer

Predicting distant metastasis (DM) in head and neck cancer (HNC) is crucial for clinical management, yet traditional machine learning methods often rely on manual tumor segmentations (regions of interest, or ROIs). These segmentations are time-consuming, subjective, and require expert knowledge. This study investigates whether medical image-based foundation models—which can extract modality-relevant features from whole-volume CT images without explicit ROI delineation—can effectively predict DM risk compared to traditional radiomics and deep learning approaches.

The researchers utilized the RADCURE dataset, consisting of 2,327 HNC patients. They compared three feature-extraction strategies: (1) handcrafted radiomics, (2) deep learning features from a Vision Transformer (ViT) trained from scratch, and (3) embeddings derived from a pre-trained CT Foundation model. These feature sets were fed into a multi-layer perceptron (MLP) to predict DM risk. The study also evaluated these features using support vector machine (SVM) classifiers and compared them against clinical-only baseline models.

The CT Foundation-based model achieved an AUC of 0.791, outperforming both the radiomics model (AUC 0.772) and the ViT model (AUC 0.753). Notably, the foundation model's performance was comparable to a combined model that integrated both radiomics and deep learning features (AUC 0.794). These results suggest that foundation models can achieve high predictive accuracy while bypassing the labor-intensive requirement of manual ROI contouring.

By removing the dependency on expert-annotated segmentations, foundation models offer a more scalable and accessible path for clinical prognostic modeling. This approach reduces the computational burden of 3D image processing and minimizes user-dependent bias, potentially accelerating the deployment of AI-driven risk assessment tools in clinical workflows.