Research map / Large language models / Reasoning and chain of thought
Learning Self-Critiquing Mechanisms for Region-Guided Chest X-Ray Report Generation
Sixing Yan, Ziao Wang, Kejing Yin, William K. Cheung, Ka Chun Cheung, Simon See
Read the paperCheck your idea against 137,217 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 a self-critiquing learning mechanism for region-guided chest X-ray report generation so that models ground diagnoses in abnormality regions rather than superficial image-report correlations.
Problem
Automatic radiology report generation from chest X-rays with supervised learning tends to learn superficial statistical correlations between images and reports and lacks grounding of findings in abnormality locations.
Approach
Introduces a self-critiquing mechanism that lets the model reason about and critique the relevant image regions when generating reports, so that abnormality grounding is learned rather than inferred from superficial correlations. Only the abstract was available, so finer details are not stated.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add all three self-critiquing mechanisms (train and test) (ablation) | RadSCR · MIMIC-CXR | CheXbert F-1: 0.561 -> 0.610 | Improved |
| add self-critiquing at inference (ablation) | RadSCR · MIMIC-CXR | CE-Abn: 0.379 -> 0.572 | Improved |
| add alternative-abnormality critique C(Abn) (ablation) | RadSCR · MIMIC-CXR | CE-Abn: 0.523 -> 0.572 | Improved |
| add false-negative critique C(Neg) (ablation) | RadSCR · MIMIC-CXR | CheXbert F-1: 0.602 -> 0.610 | Improved |
| add alternative-image critique C(Img) (ablation) | RadSCR · MIMIC-CXR | RadGraph-F1 (Complete): 0.371 -> 0.367 | No clear effect |
| add LLM decoder (ablation) | RadSCR · MIMIC-CXR | RadNLI F1: 0.326 -> 0.408 | Improved |
| add LLM decoder (ablation) | RadSCR · MIMIC-CXR | CheXbert F-1: 0.611 -> 0.610 | No clear effect |
| add abnormality prototypes (ablation) | RadSCR · MIMIC-CXR | CE-Abn: 0.377 -> 0.572 | Improved |
| add abnormality prototypes (ablation) | RadSCR · MIMIC-CXR | CE-Organ: 0.751 -> 0.744 | Mixed |
| increase sampled alternatives to Np=2 | RadSCR · MIMIC-CXR | clinical metrics | Improved |
| sample only hard alternatives instead of random | RadSCR · MIMIC-CXR | CheXbert F-1: 0.610 -> 0.577 | Hurt |
| use region-aware RadSCR instead of MLP head | Swin · MIMIC-CXR abnormality prediction | ROC-AUC: 0.756 -> 0.900 | Improved |
Closest papers by meaning
- Instance-level Expert Knowledge and Aggregate Discriminative Attention for Radiology Report Generation (CVPR 2024 · similarity 0.87)
- CXReasonBench: A Benchmark for Evaluating Structured Diagnostic Reasoning in Chest X-rays (NeurIPS 2025 · similarity 0.86)
- Reporting Practice Matters: The Impact of Reference Choice on Chest X-ray Report Evaluation (NeurIPS 2026 · similarity 0.86)
- DART: Disease-aware Image-Text Alignment and Self-correcting Re-alignment for Trustworthy Radiology Report Generation (CVPR 2025 · similarity 0.85)
- Bootstrapping Large Language Models for Radiology Report Generation (AAAI 2024 · similarity 0.85)
- Bootstrapping Chest CT Image Understanding by Distilling Knowledge from X-ray Expert Models (CVPR 2024 · similarity 0.85)
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.
