Toward Training-Free Zero-Shot Anomaly Detection in 3D Medical Images: A Batch-Based Approach Using 2D Foundation Models

A team of researchers has proposed a training-free framework that applies 2D vision foundation models to zero-shot anomaly detection in 3D medical scans, addressing a gap left by methods designed for flat images. The framework, called CS3F, was detailed in a paper submitted to arXiv on June 17, 2026 [1]. Zero-shot anomaly detection is appealing for clinical imaging because systems must cope with varying acquisition protocols, shifting patient populations, and rare pathologies that lack annotated training data [2]. Most existing zero-shot methods are built for 2D images, and extending them to 3D volumes is hampered by a scarcity of large-scale volumetric foundation models and the challenge of using full volumetric context [2]. CS3F works by decomposing each 3D volume along multiple anatomical axes and encoding the resulting slices with a frozen 2D vision transformer [2]. Neighboring slice features are pooled into localized volumetric tokens. The system then calculates anomaly scores through cross-subject mutual similarity: tokens that find no close match in other subjects receive higher scores, flagging potential lesions [2]. To counteract the signal loss that depth pooling can cause for small focal anomalies, the authors introduced a coarse-to-fine tokenization strategy that enables fine-resolution scoring without exhaustive pairwise matching [2]. The method was tested on brain MRI scans covering metastases, glioma, and stroke, and was further validated on lung CT scans to probe performance beyond atlas-aligned brain imaging [2]. The results indicate that frozen 2D foundation models can support anomaly localization in 3D medical images, though the benefit of fine tokenization depends heavily on lesion contrast and the imaging modality used [2]. The work lands at a moment when the machine-learning community is grappling with the cost and difficulty of producing high-quality labeled training datasets for supervised learning [3]. Unlabeled datasets for unsupervised tasks also remain expensive to assemble, which heightens interest in approaches that bypass task-specific training altogether [3]. The CS3F framework does not require fine-tuning or volumetric foundation models, relying instead on batch-level comparisons across subjects [2]. The paper’s supporting code and media links point to platforms including Hugging Face, DagsHub, and ScienceCast, signaling an intent to release artifacts for community use [1]. The authors have not yet published peer-reviewed results, and the preprint has not been independently replicated. The framework’s reliance on cross-subject similarity means its sensitivity may vary with cohort composition and the rarity of a given anomaly type, a limitation the paper acknowledges in its modality-dependent findings [2].