Classification on ChestX-ray14 (test)
0.8361AUROCRad-DINO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Rad-DINOPre-train Data=LVD (142M) & X-rays (public+private), Backbone=ViT-B2025.04 | 0.8361 | — | |
| CheXWorldPre-train Data=X-rays (0.5M), Backbone=ViT-B2025.04 | 0.8358 | — | |
| Adam-v2Pre-train Data=IN-21k (12M) & X-rays (0.9M), Backbone=ConvNeXt-B2025.04 | 0.834 | — | |
| SimMIMPre-train Data=IN-21k (12M) & X-rays (0.9M), Backbone=Swin-B2025.04 | 0.8304 | — | |
| MAEPre-train Data=X-rays (0.5M), Backbone=ViT-B2025.04 | 0.83 | — | |
| ConvNeXt-BInitialization strategy=DINOv3, Resolution=512x512, Evaluation protocol=full fine-tuning2025.10 | 0.823 | — | |
| CoNNSZero-shot=true, Venue=Ours2026.05 | 0.819 | — | |
| ConvNeXt-BInitialization strategy=ImageNet, Resolution=512x512, Evaluation protocol=full fine-tuning2025.10 | 0.815 | 0.001 | |
| ViT-BInitialization strategy=DINOv3, Resolution=512x512, Evaluation protocol=full fine-tuning2025.10 | 0.813 | — | |
| CARZeroZero-shot=true, Venue=CVPR’242026.05 | 0.811 | — | |
| ConvNeXt-BInitialization strategy=DINOv3, Resolution=224x224, Evaluation protocol=full fine-tuning2025.10 | 0.804 | — | |
| RadZeroZero-shot=true, Venue=NeurIPS’252026.05 | 0.804 | — | |
| ViT-BInitialization strategy=DINOv2, Resolution=224x224, Evaluation protocol=full fine-tuning2025.10 | 0.801 | 0.006 | |
| ViT-BInitialization strategy=DINOv3, Resolution=224x224, Evaluation protocol=full fine-tuning2025.10 | 0.801 | — | |
| LVM-MedPre-train Data=Medical (1.3M), Backbone=ViT-B2025.04 | 0.8008 | — | |
| ViT-BInitialization strategy=DINOv2, Resolution=512x512, Evaluation protocol=full fine-tuning2025.10 | 0.8 | 0.006 | |
| BEITPre-train Data=IN-21k (12M), Backbone=ViT-B2025.04 | 0.7991 | — | |
| ConvNeXt-BInitialization strategy=ImageNet, Resolution=224x224, Evaluation protocol=full fine-tuning2025.10 | 0.798 | 0.001 | |
| ViT-BInitialization strategy=ImageNet, Resolution=512x512, Evaluation protocol=full fine-tuning2025.10 | 0.795 | 0.006 | |
| MoCo-v3Pre-train Data=IN-1k (1.3M), Backbone=ViT-B2025.04 | 0.792 | — | |
| ViT-BInitialization strategy=ImageNet, Resolution=224x224, Evaluation protocol=full fine-tuning2025.10 | 0.79 | 0.006 | |
| KADZero-shot=true, Venue=Nat. Commun.’232026.05 | 0.789 | — | |
| DINOPre-train Data=IN-1k (1.3M), Backbone=ViT-B2025.04 | 0.7837 | — | |
| Non-privateepsilon (ϵ)=N/A, delta (δ)=N/A, Backbone=ViT, Tuning Protocol=LoRA finetuning2025.10 | 0.7363 | — | |
| BioViL-TZero-shot=true, Venue=CVPR’232026.05 | 0.729 | — | |
| MedKLIPZero-shot=true, Venue=ICCV’232026.05 | 0.726 | — | |
| ScratchPre-train Data=-, Backbone=ViT-B2025.04 | 0.7169 | — | |
| NoiseCurveepsilon (ϵ)=8, delta (δ)=10−5, Backbone=ViT, Tuning Protocol=LoRA finetuning2025.10 | 0.6452 | — | |
| BiomedCLIPZero-shot=true, Venue=NEJM AI’242026.05 | 0.639 | — | |
| DP-BANDMFepsilon (ϵ)=8, delta (δ)=10−5, Backbone=ViT, Tuning Protocol=LoRA finetuning2025.10 | 0.6255 | — | |
| NoiseCurveepsilon (ϵ)=5, delta (δ)=10−5, Backbone=ViT, Tuning Protocol=LoRA finetuning2025.10 | 0.6255 | — | |
| DP-BANDMFepsilon (ϵ)=5, delta (δ)=10−5, Backbone=ViT, Tuning Protocol=LoRA finetuning2025.10 | 0.5932 | — | |
| DP-SGDepsilon (ϵ)=8, delta (δ)=10−5, Backbone=ViT, Tuning Protocol=LoRA finetuning2025.10 | 0.5921 | — | |
| DP-SGDepsilon (ϵ)=5, delta (δ)=10−5, Backbone=ViT, Tuning Protocol=LoRA finetuning2025.10 | 0.5816 | — |