Research map / Label-efficient and robust learning / Domain adaptation and generalization
What Makes Transfer Learning Work for Medical Images: Feature Reuse & Other Factors
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ImageNet pretrained weights rather than layer-wise weight statistics or random initialization | DeiT-S · APTOS2019, DDSM, ISIC2019, CheXpert, PatchCamelyon | task metric reported for each dataset: 0.721 $\pm$ 0.016 -> 0.894 $\pm$ 0.017; 0.895 $\pm$ 0.005 -> 0.949 $\pm$ 0.011; 0.607 $\pm$ 0.017 -> 0.824 $\pm$ 0.008; | Helped |
| Use ImageNet pretrained weights rather than layer-wise weight statistics or random initialization | SWIN-T · APTOS2019, DDSM, ISIC2019, CheXpert, PatchCamelyon | task metric reported for each dataset: 0.722 $\pm$ 0.017 -> 0.906 $\pm$ 0.005; 0.900 $\pm$ 0.004 -> 0.961 $\pm$ 0.007; 0.654 $\pm$ 0.008 -> 0.833 $\pm$ 0.008; | Helped |
| Use ImageNet pretrained weights instead of random or statistics-based initialization | InceptionV3 · APTOS2019, DDSM, ISIC2019 | task metric reported for each dataset: 0.835 $\pm$ 0.012 -> 0.873 $\pm$ 0.007; 0.923 $\pm$ 0.003 -> 0.939 $\pm$ 0.010; 0.668 $\pm$ 0.008 -> 0.758 $\pm$ 0.011 | Helped |
| Use ImageNet pretrained weights instead of statistics-based initialization | ResNet50 · APTOS2019, DDSM, ISIC2019 | task metric reported for each dataset: 0.848 $\pm$ 0.006 -> 0.888 $\pm$ 0.004; 0.933 $\pm$ 0.006 -> 0.957 $\pm$ 0.003; 0.635 $\pm$ 0.012 -> 0.795 $\pm$ 0.011 | Helped |
| Transfer ImageNet weights rather than only their statistics | DeiT-S and SWIN-T · CheXpert and PatchCamelyon | ROC-AUC: DeiT-S: 0.734 $\pm$ 0.002 -> 0.792 $\pm$ 0.001; 0.916 $\pm$ 0.005 -> 0.962 $\pm$ 0.003; SWIN-T: 0.785 $\pm$ 0.000 -> 0.8 | Helped |
| Transfer ImageNet weights rather than statistics-based initialization | ResNet50 and InceptionV3 · CheXpert and PatchCamelyon | ROC-AUC: ResNet50 CheXpert: 0.794 $\pm$ 0.001 -> 0.800 $\pm$ 0.001; PatchCamelyon: 0.959 $\pm$ 0.003 -> 0.960 $\pm$ 0.006; Incept | No clear effect |
| Use weight statistics initialization instead of random initialization | DeiT-S · APTOS2019, ISIC2019 | task metric reported for each dataset: 0.684 $\pm$ 0.017 -> 0.721 $\pm$ 0.016; 0.576 $\pm$ 0.013 -> 0.607 $\pm$ 0.017 | Helped |
| Keep the first 5 transformer blocks and discard the remaining blocks (ablation) | DeiT-S · APTOS2019, DDSM, ISIC2019, CheXpert, PatchCamelyon | task metric reported for each dataset: 0.894 $\pm$ 0.017 -> 0.894 $\pm$ 0.005; 0.949 $\pm$ 0.011 -> 0.951 $\pm$ 0.001; 0.824 $\pm$ 0.008 -> 0.812 $\pm$ 0.016; | No clear effect |
| Increase the number of transferred layers (ablation) | CNNs and vision transformers · medical image datasets | convergence speed | Helped |
| Increase model capacity (ablation) | DeiT and ResNet families · medical image datasets | transfer learning performance | Helped |
Closest papers by meaning
- Transfusion: Understanding Transfer Learning for Medical Imaging (NeurIPS 2019 · similarity 0.93)
- Knowledge Transfer Scaling Laws for 3D Medical Imaging (NeurIPS 2026 · similarity 0.87)
- How Well Do Supervised 3D Models Transfer to Medical Imaging Tasks? (ICLR 2024 · similarity 0.86)
- Do Multiple Instance Learning Models Transfer? (ICML 2025 · similarity 0.83)
- Features are fate: a theory of transfer learning in high-dimensional regression (ICML 2025 · similarity 0.83)
- Transferring Knowledge From Large Foundation Models to Small Downstream Models (ICML 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.