Research map / Label-efficient and robust learning / Self-supervised and contrastive
Preservational Learning Improves Self-Supervised Medical Image Models by Reconstructing Diverse Contexts
Hong-Yu Zhou, Chixiang Lu, Sibei Yang, Xiaoguang Han, Yizhou Yu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| add self-reconstruction branch (ablation) | ContraLoss · Chest14->Chest14 (9:1) | AUC: 71.7 -> 72.5 | Improved |
| add TransAtt (with rotation) to contrastive loss (ablation) | ContraLoss · Chest14->Chest14 (9:1) | AUC: 71.7 -> 74.4 | Improved |
| add CrossMix to contrastive loss (ablation) | ContraLoss · Chest14->Chest14 (9:1) | AUC: 71.7 -> 73.7 | Improved |
| add rotation transformations to TransAtt indicator (beyond flip) (ablation) | TransAtt · Chest14->Chest14 (9:1) | AUC: 73.5 -> 74.4 | Improved |
| replace self-reconstruction with rotation prediction (ablation) | ContraLoss + pretext · Chest14->Chest14 (9:1) | AUC: 72.5 -> 72.3 | Hurt |
| halve CrossMix alpha from 1 to 0.5 (ablation) | CrossMix · Chest14->Chest14 (9:1) | AUC: 73.7 -> 73.3 | Hurt |
| combine TransAtt and CrossMix into full PCRL (ablation) | ContraLoss · Chest14->Chest14 (8:2) | AUC: 74.8 -> 78.8 | Improved |
| add TransAtt over self-reconstruction on rotation pretext task (ablation) | ContraLoss + Self-Recons. · Chest14 rotation prediction | accuracy: 87.9 -> 91.0 | Improved |
| add TransAtt over self-reconstruction on position pretext task (ablation) | ContraLoss + Self-Recons. · Chest14 relative position prediction | accuracy: 84.5 -> 89.3 | Improved |
| add CrossMix to contrastive loss on position pretext (ablation) | ContraLoss · Chest14 relative position prediction | accuracy: 82.3 -> 88.3 | Improved |
| use PCRL instead of MoCov2 for ImageNet pretraining | self-supervised pretraining · Cityscapes | mIoU: 76.3 -> 77.3 | Improved |
| use PCRL instead of MoCov2 for ImageNet pretraining | self-supervised pretraining · COCO | mAP: 40.5 -> 41.3 | Improved |
Closest papers by meaning
- A Unified Visual Information Preservation Framework for Self-supervised Pre-training in Medical Image Analysis (TPAMI 2023 · similarity 0.88)
- Contrastive learning of global and local features for medical image segmentation with limited annotations (NeurIPS 2020 · similarity 0.86)
- Co2L: Contrastive Continual Learning (ICCV 2021 · similarity 0.86)
- When Does Contrastive Visual Representation Learning Work? (CVPR 2022 · similarity 0.86)
- Separating common from salient patterns with Contrastive Representation Learning (ICLR 2024 · similarity 0.85)
- Variational Supervised Contrastive Learning (NeurIPS 2025 · similarity 0.84)
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.
