Research map / 3D vision / 3D generation
Enhanced Contrastive Learning with Multi-view Longitudinal Data for Chest X-ray 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 MLRG, a two-stage chest X-ray report generation method that pretrains with multi-view longitudinal contrastive learning supervised by report structure and uses tokenized absence encoding to handle missing prior knowledge.
Problem
Chest X-ray report generation typically uses single or fixed-view images and ignores multi-view longitudinal data; existing longitudinal methods still use single current images, and missing patient priors (INDICATION, previous report/image) are hard to handle flexibly.
Approach
Multi-view longitudinal contrastive learning: a fusion network (cross-attention with the anchor current view as query over other current views and the previous image) produces spatiotemporal features aligned with report factual serializations via multi-positive and instance-wise contrastive losses; special [NHI]/[NHPR] tokens stand in for absent priors.
Key results
- MLRG reaches BLEU-4 0.158 on MIMIC-CXR, a 2.3% gain over the best peer method, and CE F1 0.505.
- On MIMIC-ABN, MLRG achieves F1 0.515, a 5.5% improvement over the best peer method.
- On Two-view CXR, MLRG improves F1 RadGraph by 2.7% (0.328 vs 0.301 for SEI).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add multi-view longitudinal contrastive learning (ablation) | MLRG · MIMIC-CXR | BLEU-4: 0.136 -> 0.158 | Helped |
| Add cross-modal alignment loss LG (ablation) | MLRG Stage 1 · MIMIC-CXR | BLEU-4: 0.136 -> 0.146 | Helped |
| Add multi-positive contrastive loss LMPC (ablation) | MLRG Stage 1 · MIMIC-CXR | BLEU-4: 0.146 -> 0.138 | Helped |
| Add patient-specific prior knowledge (ablation) | MLRG Stage 2 · MIMIC-CXR | BLEU-4: 0.118 -> 0.158 | Helped |
| Add indication (ablation) | MLRG · MIMIC-CXR | BLEU-4: 0.118 -> 0.153 | Helped |
| Add previous report (ablation) | MLRG · MIMIC-CXR | BLEU-4: 0.118 -> 0.118 | Helped |
| Use current multi-view images instead of single images (ablation) | MLRG · MIMIC-CXR | BLEU-4: 0.151 -> 0.158 | Helped |
| Add previous images through the MLF network (ablation) | MLRG · MIMIC-CXR | BLEU-4: 0.138 -> 0.158 | Helped |
| Include indication in the test subset | MLRG · MIMIC-CXR | F1 RadGraph: 0.254 -> 0.318 | Helped |
| Include current multi-view images in the test subset | MLRG · MIMIC-CXR | F1 RadGraph: 0.266 -> 0.301 | Helped |
| Include multi-view longitudinal data in the test subset | MLRG · MIMIC-CXR | F1 RadGraph: 0.276 -> 0.300 | Helped |
Limitations
- Lower recall than B-LLM on MIMIC-CXR
- Uses only the most recent previous visit and a single previous frontal image
- Does not model subtle per-observation accuracy gains on imbalanced labels such as pneumothorax and fracture
Experimental setup
Datasets: MIMIC-CXR MIMIC-ABN Two-view CXR IU X-ray
Benchmarks: MIMIC-CXR MIMIC-ABN Two-view CXR
Metrics: BLEU-1 BLEU-2 BLEU-3 BLEU-4 METEOR ROUGE-L CheXpert precision CheXpert recall CheXpert F1 F1 RadGraph
Baselines: R2Gen CMN SA MET KiUT CoFE MAN B-LLM DCG Med-LLM SEI FMVP
Builds on: CLIP MGCA SEI FSE CXRMate
Closest papers by meaning
- HC-LLM: Historical-Constrained Large Language Models for Radiology Report Generation (AAAI 2025 · similarity 0.88)
- Bootstrapping Chest CT Image Understanding by Distilling Knowledge from X-ray Expert Models (CVPR 2024 · similarity 0.86)
- Addressing Asynchronicity in Clinical Multimodal Fusion via Individualized Chest X-ray Generation (NeurIPS 2024 · similarity 0.86)
- Multimodal Disease Progression Modeling via Spatiotemporal Disentanglement and Multiscale Alignment (NeurIPS 2025 · similarity 0.86)
- AHIVE: Anatomy-aware Hierarchical Vision Encoding for Interactive Radiology Report Retrieval (CVPR 2024 · similarity 0.85)
- CXReasonBench: A Benchmark for Evaluating Structured Diagnostic Reasoning in Chest X-rays (NeurIPS 2025 · 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.