Research map / Label-efficient and robust learning / Recognition and long-tail
Combining Graph Transformers Based Multi-Label Active Learning and Informative Data Augmentation for Chest Xray Classification
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 DAMLAL, a multi-label active learning method that scores samples with graph multiset transformer attention over saliency-based label graphs and augments the top informative samples with VAE-generated, informativeness-filtered variants.
Problem
Active learning for multi-label chest X-ray classification is poorly served by single-label informativeness criteria, and existing augmentation-based AL methods (e.g., LADA) do not handle multiple disease labels.
Approach
Represent disease labels as nodes of a saliency-similarity graph, use Graph Multiset Transformer attention pooling to compute a multi-label informativeness score (MIS), and generate VAE variants of base informative images that are kept only if they preserve labels and add novel information (label and redundancy scores).
Key results
- On CheXpert, DAMLAL reaches 98.21 AUC with 100% data versus 96.71 for GESTALT and 95.73 for LADA.
- DAMLAL outperforms the fully supervised DenseNet-121 (90.23 AUC) at 37% of training data, while GESTALT needs 45%.
- Radiologists judged 94.1% of 1000 VAE-generated images realistic by at least one expert; MIS correlates with uncertainty at r=0.92.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DAMLAL instead of GESTALT | Chest X-ray multi-label active learning · CheXpert, 10% training data | AUC: 70.82 -> 72.24 | Helped |
| Use DAMLAL instead of GESTALT | Chest X-ray multi-label active learning · CheXpert, 100% training data | AUC: 96.71 -> 98.21 | Helped |
| Use GradCAM saliency maps instead of Deep Taylor (ablation) | DAMLAL · CheXpert, 10% training data | AUC: 72.24 -> 71.63 | Hurt |
| Add informative augmentation instead of conventional augmentation (ablation) | DAMLAL · CheXpert, 10% training data | AUC: 71.23 -> 72.24 | Helped |
| Use graph transformers instead of summing graph weights (ablation) | DAMLAL · CheXpert, 10% training data | AUC: 71.01 -> 72.24 | Helped |
| Add the label-preservation score (ablation) | DAMLAL informative augmentation · CheXpert, 10% training data | AUC: 68.22 -> 72.24 | Helped |
| Add the redundancy-avoidance score (ablation) | DAMLAL informative augmentation · CheXpert, 10% training data | AUC: 65.61 -> 72.24 | Helped |
| Use multiset pooling instead of conventional pooling (ablation) | DAMLAL graph representation · CheXpert, 10% training data | AUC: 64.13 -> 72.24 | Helped |
| Add graph attention (ablation) | DAMLAL graph representation · CheXpert, 10% training data | AUC: 62.21 -> 72.24 | Helped |
| Add self-attention (ablation) | DAMLAL graph representation · CheXpert, 10% training data | AUC: 61.35 -> 72.24 | Helped |
| Add informative augmentation instead of conventional augmentation (ablation) | DAMLAL · CheXpert, 100% training data | AUC: 97.01 -> 98.21 | Helped |
| Use graph transformers instead of summing graph weights (ablation) | DAMLAL · CheXpert, 100% training data | AUC: 96.81 -> 98.21 | Helped |
Surprising findings
- Random sample selection did not outperform the fully supervised reference at low data fractions, unlike other AL methods.
- Removing the redundancy score hurt performance more than removing the label score.
Limitations
- Synthetic VAE images are judged realistic for only 91-94% of samples, with 5.9% judged unrealistic by all experts.
- Hyperparameters (eta, gamma, lambda) were tuned on the evaluation setting, and sensitivity analysis shows substantial performance dependence on them.
- Performance degrades significantly with Gaussian noise above sigma=0.01.
Experimental setup
Datasets: CheXpert NIH ChestXray14
Benchmarks: CheXpert NIH ChestXray14
Metrics: AUC p-value training time inference time radiologist realism agreement
Baselines: Random sampling Uncertainty sampling LEMAL CVIRS AlphaMix GAL LADA GESTALT FSL
Builds on: Graph Multiset Transformer LADA GESTALT Variational Autoencoder
Closest papers by meaning
- ACPL: Anti-Curriculum Pseudo-Labelling for Semi-Supervised Medical Image Classification (CVPR 2022 · similarity 0.84)
- Novel Class Discovery in Chest X-rays via Paired Images and Text (AAAI 2024 · similarity 0.83)
- Many-to-One Distribution Learning and K-Nearest Neighbor Smoothing for Thoracic Disease Identification (AAAI 2021 · similarity 0.83)
- Enhanced Contrastive Learning with Multi-view Longitudinal Data for Chest X-ray Report Generation (CVPR 2025 · similarity 0.82)
- CARZero: Cross-Attention Alignment for Radiology Zero-Shot Classification (CVPR 2024 · similarity 0.81)
- Towards Cost-Effective Learning: A Synergy of Semi-Supervised and Active Learning (CVPR 2025 · similarity 0.81)
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.