DINOv3-MIL detects kidney tumors and cysts with 80% accuracy

A new computer vision paper, DINOv3-MIL, demonstrates that foundation models like DINOv3 ViT-H/16, trained on natural images, can be effectively transferred to volumetric medical image analysis without domain-specific pre-training. The key challenge is aggregating tens of thousands of patch tokens per 3D kidney study (55,296 tokens). The authors compare three aggregator architectures: a CLS-token linear probe, gated attention multiple instance learning (MIL), and a prototype head (ProtoViT). On the KiTS23 dataset (966 kidneys, 97 test), attention MIL achieves the highest performance with AUROC 0.74 for tumor detection and 0.80 for cyst detection. Importantly, the attention mechanism is highly localized, showing 7.5-9.8x enrichment within annotated lesions, meaning the model focuses precisely on relevant tissue.

The prototype head, while more interpretable in principle, completely fails on cyst detection (AUROC 0.51), revealing a clear interpretability-performance trade-off at this token scale. The CLS-token linear probe performs in between. These results, accepted as a poster at MIUA 2026 and to appear in Frontiers in Medical Technology, suggest that gated attention MIL is currently the best choice for aggregating foundation-model patch tokens in volumetric medical tasks. The method enables per-kidney multi-label detection without needing full 3D segmentation, potentially reducing annotation effort in clinical workflows. The paper's code and data are linked on arXiv.