Research map / Label-efficient and robust learning / Self-supervised and contrastive
AFiRe: Anatomy-Driven Self-Supervised Learning for Fine-Grained Representation in Radiographic Images
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AFiRe, a self-supervised ViT pre-training framework for chest X-rays that combines token-wise anatomy-guided contrastive learning (with spatial-aware prototypes) and pixel-level anomaly-removal restoration, using Synthetic Lesion Masks as augmentation.
Problem
Standard self-supervised contrastive methods focus on global image-level discrimination and miss fine-grained anatomical details needed for radiographic analysis, while existing medical SSL methods lack anatomical semantics or require region pre-selection.
Approach
Treats each ViT token as an independent sample aligned to spatial-aware prototypes (one per token position), contrasting token probability distributions for structure and category consistency, and combines this with restoring normal anatomy by replacing synthetic-lesion tokens with learnable mask tokens; lesions are generated via Synthetic Lesion Masks that improve on AnatPaste.
Key results
- On NIH with 1% labels, AFiRe reaches 63.2% AUC vs 62.9% for the best radiography-specific baseline PCRL; on SIIM-ACR at 100% labels it reaches 92.4% Dice vs 83.9% for DiRA.
- On ChildCXRs anomaly detection, AFiRe obtains 90.8% AUC and 83.9% ACC, while AnatPaste reports 91.4% AUC and 83.0% ACC.
- Ablation on ChildCXRs linear probing: full AFiRe reaches 89.4% AUC, 83.8% ACC, and 84.5% F1, and removing the restoration loss drops AUC to 86.2%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use AFiRe instead of Adamv2 | ViT-B transfer learning · NIH 5-C | AUC: 84.1 -> 84.5 | Helped |
| use AFiRe instead of Adamv2 | ViT-B transfer learning · CXP 5-C | AUC: 87.5 -> 89.0 | Helped |
| use AFiRe instead of PCRL | ViT-B transfer learning · SIIM-ACR 5-C | Dice: 80.4 -> 87.7 | Helped |
| replace SLM with AnatPaste (ablation) | AFiRe · ChildCXR linear probing | AUC: 89.4 -> 88.9 | Hurt |
| add spatial-aware prototypes (ablation) | AFiRe · ChildCXR linear probing | AUC: 83.1 -> 89.4 | Helped |
| add category-consistency contrastive loss (ablation) | AFiRe · ChildCXR linear probing | AUC: 87.7 -> 89.4 | Helped |
| add structure-consistency contrastive loss (ablation) | AFiRe · ChildCXR linear probing | AUC: 84.5 -> 89.4 | Helped |
| add both category- and structure-consistency contrastive losses (ablation) | AFiRe · ChildCXR linear probing | AUC: 76.9 -> 89.4 | Helped |
| add anomaly-removal pixel restoration (ablation) | AFiRe · ChildCXR linear probing | AUC: 86.2 -> 89.4 | Helped |
| use probability contrast instead of feature contrast (ablation) | AFiRe · ChildCXR linear probing | AUC: 84.8 -> 89.4 | Helped |
| add trainable z_mask tokens to replace abnormal tokens (ablation) | AFiRe · ChildCXR linear probing | AUC: 85.7 -> 89.4 | Helped |
| add consistency between abnormal tokens (ablation) | AFiRe · ChildCXR linear probing | AUC: 89.4 -> 77.9 | Hurt |
Surprising findings
- Enforcing consistency among abnormal tokens (the I(Q-,Q-) term) reduced performance sharply, with AUC falling to 77.9%.
Limitations
- Tends to prioritize primary abnormalities and may overlook larger affected regions, as stated by the authors.
- Generated images show damage where lesion density covers anatomy, lowering density in normal chest regions.
Experimental setup
Datasets: MIMIC-CXR-JPG ChildCXRs NIH ChestX-ray14 CXP (CheXpert) SIIM-ACR
Benchmarks: NIH ChestX-ray14 CXP SIIM-ACR ChildCXRs
Metrics: AUC Dice ACC F1
Baselines: ViT-B (ImageNet supervised) MoCov2 DINO MAE C2L TransVW MG DiRA PCRL Adamv1 Adamv2 MemAE
Builds on: DINO SwAV AnatPaste Adamv2 SelfPatch
Closest papers by meaning
- Large-scale and Fine-grained Vision-language Pre-training for Enhanced CT Image Understanding (ICLR 2025 · similarity 0.85)
- Autoregressive Sequence Modeling for 3D Medical Image Representation (AAAI 2025 · similarity 0.83)
- Modality-Agnostic Structural Image Representation Learning for Deformable Multi-Modality Medical Image Registration (CVPR 2024 · similarity 0.83)
- CheXWorld: Exploring Image World Modeling for Radiograph Representation Learning (CVPR 2025 · similarity 0.83)
- S2D-Align: Shallow-to-Deep Auxiliary Learning for Anatomically-Grounded Radiology Report Generation (AAAI 2026 · similarity 0.82)
- Representing Part-Whole Hierarchies in Foundation Models by Learning Localizability Composability and Decomposability from Anatomy via Self Supervision (CVPR 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.