Research map / Label-efficient and robust learning / Recognition and long-tail
Transfusion: Understanding Transfer Learning for Medical Imaging
Read the paperCheck 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 |
|---|---|---|---|
| Initialize Resnet-50 with ImageNet pretrained weights instead of random weights | Resnet-50 · Retina | AUC-ROC: $96.4\% \pm 0.05$ -> $96.7\% \pm 0.04$ | No clear effect |
| Initialize Inception-v3 with ImageNet pretrained weights instead of random weights | Inception-v3 · Retina | AUC-ROC: $96.6\% \pm 0.13$ -> $96.7\% \pm 0.05$ | No clear effect |
| Use ImageNet transfer instead of random initialization | Resnet-50 · CheXpert Atelectasis | AUC: 79.52$\pm$0.31 -> 79.76$\pm$0.47 | Helped |
| Use ImageNet transfer instead of random initialization | Resnet-50 · CheXpert Cardiomegaly | AUC: 75.23$\pm$0.35 -> 74.93$\pm$1.41 | Hurt |
| Use ImageNet transfer instead of random initialization | Resnet-50 · CheXpert Consolidation | AUC: 85.49$\pm$1.32 -> 84.42$\pm$0.65 | Hurt |
| Use ImageNet transfer instead of random initialization | Resnet-50 · CheXpert Edema | AUC: 88.34$\pm$1.17 -> 88.89$\pm$1.66 | Helped |
| Use ImageNet transfer instead of random initialization | Resnet-50 · CheXpert Pleural Effusion | AUC: 88.70$\pm$0.13 -> 88.07$\pm$1.23 | Hurt |
| Use ImageNet pretrained weights rather than random initialization | Resnet50 · Retina, 5k datapoints | AUC-ROC: $92.2 \%$ -> $94.6 \%$ | Helped |
| Use ImageNet pretrained weights rather than random initialization | CBR-LargeT · Retina, 5k datapoints | AUC-ROC: $93.6 \%$ -> $93.9 \%$ | No clear effect |
| Use ImageNet pretrained weights rather than random initialization | CBR-LargeW · Retina, 5k datapoints | AUC-ROC: $93.6 \%$ -> $93.7 \%$ | No clear effect |
| Preserve per-layer pretrained weight mean and variance while destroying pretrained features (ablation) | Mean Var initialization · Retina and CheXpert | convergence speed | Helped |
| Reuse pretrained weights in the lowest layers, especially conv1 (ablation) | Resnet-50 weight transfusion · Retina | convergence speed | Helped |
Closest papers by meaning
- What Makes Transfer Learning Work for Medical Images: Feature Reuse & Other Factors (CVPR 2022 · similarity 0.93)
- Knowledge Transfer Scaling Laws for 3D Medical Imaging (NeurIPS 2026 · similarity 0.89)
- What is being transferred in transfer learning? (NeurIPS 2020 · similarity 0.88)
- How Well Do Supervised 3D Models Transfer to Medical Imaging Tasks? (ICLR 2024 · similarity 0.88)
- Do Better ImageNet Models Transfer Better? (CVPR 2019 · similarity 0.88)
- Features are fate: a theory of transfer learning in high-dimensional regression (ICML 2025 · similarity 0.85)
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.