Toward Training-Free Zero-Shot Anomaly Detection in 3D Medical Images: A Batch-Based Approach Using 2D Foundation Models
Researchers have developed CS3F, a novel framework for training-free zero-shot anomaly detection in 3D medical images. This approach utilizes existing 2D foundation models by decomposing 3D volumes into slices and encoding them with a 2D vision transformer. Anomaly scores are then derived from the similarity of these encoded features across different subjects, identifying tokens that deviate significantly from the norm. The method has been evaluated on brain MRI scans for conditions like metastases, glioma, and stroke, and further validated on lung CT scans to assess its generalizability. AI
IMPACT Enables anomaly detection in 3D medical imaging without specific training data, potentially improving diagnostic capabilities.