Medical Image Classification on ChestMNIST
94.8AccuracyResNet-50
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| ResNet-50Resolution=2242026.02 | 94.8 | 0.773 | |
| Google AutoML Vision2026.02 | 94.8 | 0.778 | |
| ViT-BCategory=Supervised (Full-Tuning)2026.03 | 94.8 | — | |
| Swin-BCategory=Supervised (Full-Tuning)2026.03 | 94.8 | — | |
| ConvNeXt-BCategory=Supervised (Full-Tuning)2026.03 | 94.8 | — | |
| SigLIP2-BCategory=Supervised (Full-Tuning)2026.03 | 94.8 | — | |
| ResNet-18Resolution=282026.02 | 94.7 | 0.768 | |
| ResNet-18Resolution=2242026.02 | 94.7 | 0.773 | |
| ResNet-50Resolution=282026.02 | 94.7 | 0.769 | |
| G-LoG bi-filtration (MLP)Sigma (σ)=02026.02 | 94.7 | 0.606 | |
| G-LoG bi-filtration (MLP)Sigma (σ)=0.52026.02 | 94.7 | 0.608 | |
| G-LoG bi-filtration (MLP)Sigma (σ)=12026.02 | 94.7 | 0.61 | |
| G-LoG bi-filtration (MLP)Sigma (σ)=1.52026.02 | 94.7 | 0.603 | |
| ResNet50Category=Supervised (Full-Tuning)2026.03 | 94.7 | — | |
| AutoKeras2026.02 | 93.7 | 0.742 | |
| Gemini-3-flashCategory=General MLLMs2026.03 | 91.6 | — | |
| GPT-5-miniCategory=General MLLMs2026.03 | 91.4 | — | |
| Clinician Mimetic WorkflowCategory=Proposed2026.03 | 90.9 | — | |
| auto-sklearn2026.02 | 77.9 | 0.649 | |
| LLaVA-MedCategory=Medical MLLMs2026.03 | 72.8 | — | |
| MedGemmaCategory=Medical MLLMs2026.03 | 60.9 | — | |
| Topo-Med (MLP)2026.02 | 53 | 0.787 | |
| Gaussian noisesigma=3.0, Backbone=ResNet-18 layer-4, Feature dimension (d)=512, Evaluation protocol=post-hoc frozen features, Head=linear-probe2026.05 | — | 0.722 | |
| KNT (full, 2-layer)Backbone=ResNet-18 layer-4, Feature dimension (d)=512, Evaluation protocol=post-hoc frozen features, Head=linear-probe, Layers=2-layer2026.05 | — | 0.749 | |
| KNT w/o key (non-secret W)Backbone=ResNet-18 layer-4, Feature dimension (d)=512, Evaluation protocol=post-hoc frozen features, Head=linear-probe2026.05 | — | 0.747 | |
| KNT w/o permutation (ReLU + key)Backbone=ResNet-18 layer-4, Feature dimension (d)=512, Evaluation protocol=post-hoc frozen features, Head=linear-probe2026.05 | — | 0.749 | |
| KNT w/o ReLU (linear, keyed)Backbone=ResNet-18 layer-4, Feature dimension (d)=512, Evaluation protocol=post-hoc frozen features, Head=linear-probe, Layers=linear2026.05 | — | 0.754 | |
| NoPeekBackbone=ResNet-18 layer-4, Feature dimension (d)=512, Evaluation protocol=post-hoc frozen features, Head=linear-probe2026.05 | — | 0.723 | |
| Raw features (no defense)Backbone=ResNet-18 layer-4, Feature dimension (d)=512, Evaluation protocol=post-hoc frozen features, Head=linear-probe2026.05 | — | 0.759 | |
| Spatial permutation (keyed)Backbone=ResNet-18 layer-4, Feature dimension (d)=512, Evaluation protocol=post-hoc frozen features, Head=linear-probe2026.05 | — | 0.759 |