Research map / Image generation / Text-to-image diffusion
DART: Disease-aware Image-Text Alignment and Self-correcting Re-alignment for Trustworthy Radiology Report Generation
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 DART, a two-stage radiology report generation framework that retrieves disease-matched reports via contrastive image-text alignment and refines generated reports with a self-correction re-alignment module.
Problem
Radiology report generation from chest X-rays often fails to capture disease-relevant findings, and retrieved reports used as prior knowledge may not match the disease findings in the input image.
Approach
Combines CLIP-style image-text contrastive learning with a disease classifier and a disease-matching constraint for image-to-text retrieval, then applies a second-stage self-correction module that re-aligns the generated report's embedding with the image embedding.
Key results
- On MIMIC-CXR, DART reports BLEU-1 0.437, BLEU-4 0.137, ROUGE-L 0.310, surpassing prior methods in these metrics.
- On MIMIC-CXR clinical efficacy, DART achieves F1 0.533, precision 0.520, recall 0.546, above I3 + C2FD (F1 0.473).
- On IU X-ray, DART reaches BLEU-4 0.208 and ROUGE-L 0.411, the best among compared methods on these metrics; ablation shows self-correction adds gains on MIMIC-CXR (BLEU-4 0.129 to 0.137).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DART instead of I3 + C2FD | Radiology report generation · MIMIC-CXR | BLEU-4: 0.128 -> 0.137 | Helped |
| Use DART instead of I3 + C2FD | Radiology report generation · MIMIC-CXR | ROUGE-L: 0.291 -> 0.310 | Helped |
| Use DART instead of I3 + C2FD | Radiology report generation · IU X-ray | BLEU-4: 0.184 -> 0.208 | Helped |
| Use DART instead of I3 + C2FD | Radiology report generation · IU X-ray | ROUGE-L: 0.390 -> 0.411 | Helped |
| Add contrastive loss (ablation) | DART report generator · MIMIC-CXR | BLEU-4: 0.111 -> 0.113 | Helped |
| Add image-to-text retrieval (ablation) | DART report generator · MIMIC-CXR | BLEU-4: 0.113 -> 0.121 | Helped |
| Add disease-matching constraint (ablation) | DART report generator · MIMIC-CXR | BLEU-4: 0.121 -> 0.129 | Helped |
| Add self-correction module (ablation) | DART report generator · MIMIC-CXR | BLEU-4: 0.129 -> 0.137 | Helped |
Surprising findings
- GPT-4 refinement of the generated report did not address omissions of key findings, while the learned self-correction module did (qualitative comparison).
Limitations
- Evaluated only on two chest X-ray benchmarks; generalization to other imaging modalities or datasets is not shown.
- Qualitative comparison with GPT-4 correction is limited to one sample; no human evaluation reported.
- Some reported baselines lack full metric coverage (dashes in tables), limiting complete comparison.
Experimental setup
Datasets: MIMIC-CXR IU X-ray
Benchmarks: MIMIC-CXR IU X-ray
Metrics: BLEU-1 BLEU-2 BLEU-3 BLEU-4 ROUGE-L METEOR CE F1 CE precision CE recall
Baselines: R2Gen R2GenCMN PPKED CMCL DCL RGRG METransformer ECRG Med-LLM MA I3 + C2FD SentSAT+KG
Builds on: CLIP R2Gen METransformer
Closest papers by meaning
- Automatic Radiology Reports Generation via Memory Alignment Network (AAAI 2024 · similarity 0.90)
- S2D-Align: Shallow-to-Deep Auxiliary Learning for Anatomically-Grounded Radiology Report Generation (AAAI 2026 · similarity 0.89)
- Image-aware Evaluation of Generated Medical Reports (NeurIPS 2024 · similarity 0.89)
- DAMPER: A Dual-Stage Medical Report Generation Framework with Coarse-Grained MeSH Alignment and Fine-Grained Hypergraph Matching (AAAI 2025 · similarity 0.88)
- AHIVE: Anatomy-aware Hierarchical Vision Encoding for Interactive Radiology Report Retrieval (CVPR 2024 · similarity 0.86)
- Bootstrapping Large Language Models for Radiology Report Generation (AAAI 2024 · similarity 0.86)
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.