Medical Image Classification on OrganMNIST3D
100AccuracyCoralBayU96H (FT)
Evaluation results
| Method | Links | ||
|---|---|---|---|
| CoralBayU96H (FT)# pre-trained data=11k, Evaluation Protocol=Fine-tuned2026.06 | 100 | — | |
| SwinUNETR# pre-trained data=—, Evaluation Protocol=Fine-tuned2026.06 | 99 | — | |
| CoralBayU96B# pre-trained data=11k, Evaluation Protocol=Linear Probing2026.06 | 99 | — | |
| CoralBayU96H# pre-trained data=11k, Evaluation Protocol=Linear Probing2026.06 | 99 | — | |
| SuPreM# pre-trained data=2.1k, Evaluation Protocol=Linear Probing2026.06 | 98 | — | |
| Universal Model# pre-trained data=2.1k, Evaluation Protocol=Linear Probing2026.06 | 96 | — | |
| VoCo-B# pre-trained data=160k, Evaluation Protocol=Linear Probing2026.06 | 96 | — | |
| AC-Tiny# of trainable params=9.5M2026.02 | 95.7 | 99.9 | |
| AC-Base# of trainable params=427M2026.02 | 95.7 | 99.9 | |
| AC-Small# of trainable params=216M2026.02 | 95.2 | 99.8 | |
| LMTTM-VMI# of trainable params=21M2026.02 | 94.8 | 99.7 | |
| Medformer2026.02 | 93.7 | 99.1 | |
| VoCo-H# pre-trained data=160k, Evaluation Protocol=Linear Probing2026.06 | 92 | — | |
| ResNet-18+3D2026.02 | 90.7 | 99.6 | |
| ResNet-18 + 3DParams=33M2026.05 | 90.7 | 99.6 | |
| Official# of trainable params=33M2026.02 | 90.7 | 99.6 | |
| Transfer-AwarePretraining Strategy=Data mixture (T = 100K)2026.05 | 90.3 | 99.4 | |
| ResNet-18+ACS2026.02 | 90 | 99.4 | |
| ResNet-18 + ACSParams=11M2026.05 | 90 | 99.4 | |
| ResNet-50+ACS2026.02 | 88.9 | 99.4 | |
| Data-ProportionalPretraining Strategy=Data mixture (T = 100K)2026.05 | 88.9 | 99.4 | |
| BSDA# of trainable params=33M2026.02 | 88.7 | 99.4 | |
| Single-Modality onlyPretraining Strategy=No mixture2026.05 | 88.6 | 99.3 | |
| ResNet-50+3D2026.02 | 88.3 | 99.4 | |
| ILPOResNet-50Params=38K2026.05 | 87.9 | 99.2 | |
| EquiLoPO (local train.)Params=418K2026.05 | 86.6 | 99.1 | |
| From scratchPretraining Strategy=No mixture2026.05 | 84.6 | 98.8 | |
| auto-sklearn2026.02 | 81.4 | 97.7 | |
| AutoKeras2026.02 | 80.4 | 97.9 | |
| ResNet-18+2.5D2026.02 | 78.8 | 97.7 | |
| ResNet-50+2.5D2026.02 | 76.9 | 97.4 | |
| SE3MovFrNet2026.05 | 74.5 | — | |
| VLLM + S2COPE (Ours)Labels=Not Required, Architecture=Qwen3-VL-8B, Evaluation Protocol=Concept-based linear probing2026.06 | 72.05 | — | |
| G-LoG bi-filtration (MLP)sigma=0.52026.02 | 70.2 | 96.1 | |
| Regular SE(3) convParams=172K2026.05 | 69.8 | — | |
| G-LoG bi-filtration (MLP)sigma=02026.02 | 68.4 | 95.8 | |
| G-LoG bi-filtration (MLP)sigma=12026.02 | 65.6 | 95.3 | |
| LF-CBMLabels=Required, Architecture=Qwen3-VL-8B, Evaluation Protocol=Concept-based linear probing2026.06 | 65.22 | — | |
| G-LoG bi-filtration (MLP)sigma=1.52026.02 | 63.4 | 94.1 | |
| VLLM (Unoptimized)Labels=Not Required, Architecture=Qwen3-VL-8B, Evaluation Protocol=Concept-based linear probing2026.06 | 58.39 | — | |
| Topo-Med (MLP)2026.02 | 55.4 | 83.7 | |
| LaBoLabels=Required, Architecture=Qwen3-VL-8B, Evaluation Protocol=Concept-based linear probing2026.06 | 22.36 | — | |
| DCLIPLabels=Required, Architecture=Qwen3-VL-8B, Evaluation Protocol=Concept-based linear probing2026.06 | 14.29 | — | |
| TextSpanLabels=Required, Architecture=Qwen3-VL-8B, Evaluation Protocol=Concept-based linear probing2026.06 | 12.42 | — |