Research map / Label-efficient and robust learning / Self-supervised and contrastive
Representing Part-Whole Hierarchies in Foundation Models by Learning Localizability Composability and Decomposability from Anatomy via Self Supervision
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.
Adam-v2 is a self-supervised pretraining framework for medical chest X-ray and fundus images that adds composability and decomposability losses to a localizability (self-distillation) objective to encode part-whole anatomical hierarchies.
Problem
Standard SSL and medical foundation models do not explicitly encode part-whole hierarchies of anatomical structures, which are a prominent property of medical images.
Approach
Extends Adam with three branches over a student/teacher encoder: localizability (multi-scale self-distillation over anchor patches), composability (aggregated part embeddings predict the whole's embedding), and decomposability (whole embedding is decoded into part embeddings), trained coarse-to-fine over anatomical scales.
Key results
- In 3-shot JSRT clavicle segmentation, Adam-v2 reaches 73.59 Dice vs 56.87 for LVM-Med, the strongest large-scale baseline.
- Large-scale Adam-v2 (ConvNeXt-B, about 1M images) achieves 83.4 mean AUC on ChestX-ray14, versus 83.0 for Medical MAE and 82.0 for LVM-Med.
- In full fine-tuning, Adam-v2 significantly outperforms all SSL counterparts (p < 0.05) across the evaluated tasks, per Fig. 7.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add composability learning to localizability (ablation) | Adam–v2 · Two downstream tasks (Fig. 8) | downstream performance | Helped |
| Add decomposability learning to localizability and composability (ablation) | Adam–v2 · Two downstream tasks (Fig. 8) | downstream performance | Helped |
| Increase anatomical granularity from m = 0 to m = 2 (ablation) | Adam–v2 · Downstream tasks (Fig. 8) | downstream performance | Helped |
| Increase anatomical granularity from m = 2 to m = 3 (ablation) | Adam–v2 · Downstream tasks (Fig. 8) | downstream performance | No clear effect |
Limitations
- Evaluation focuses on chest X-ray and fundus images, so generality to other anatomy is not shown.
- Compute resources for pretraining are not stated.
Experimental setup
Datasets: ChestX-ray14 JSRT VinDR-Rib ChestX-Det SIIM-ACR VinDr-CXR NIH Shenzhen DRIVE Drishti-GS EyePACS
Benchmarks: JSRT SIIM-ACR ChestX-Det ChestX-ray14 DRIVE Drishti-GS VinDR-Rib VinDr-CXR NIH Shenzhen
Metrics: Dice IoU AUC cosine similarity t-test significance
Baselines: RadImageNet LVM-Med DINO DenseCL DiRA TransVW PCRL Medical-MAE MoCo-v2
Builds on: Adam DINO GLOM
Closest papers by meaning
- Anatomy-Preserving Unpaired Medical Image Translation via Shared Latent Anchoring (NeurIPS 2026 · similarity 0.87)
- Anatomical Knowledge Mining and Matching for Semi-supervised Medical Multi-structure Detection (AAAI 2025 · similarity 0.85)
- GRAPE: Graph-Augmented Prototype Explanations for Interactive Medical Image Diagnosis (NeurIPS 2026 · similarity 0.84)
- Unsupervised Domain Adaptation for Anatomical Structure Detection in Ultrasound Images (ICML 2024 · similarity 0.84)
- Few-Shot Learning from Gigapixel Images via Hierarchical Vision-Language Alignment and Modeling (NeurIPS 2025 · similarity 0.84)
- Implicit Modeling of Non-rigid Objects with Cross-Category Signals (AAAI 2024 · similarity 0.83)
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.