Research map / Vision-language and multimodal / Multimodal LLMs
Mitigating Entity Hallucinations in 3D Radiology Report Generation via Dual-Stream Alignment
Lingyu Zhou, Yue Yu, Zhang Yi, Xiuyuan Xu
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 dual-stream alignment approach to reduce entity hallucinations in 3D CT radiology report generation by using medical entity phrases as multi-modal prompts.
Problem
Entity hallucination in radiology report generation for 3D CT scans, where complex spatial context amplifies factual errors; existing methods use unified cross-attention for volume-phrase alignment, which fails to account for anatomical structure differences.
Approach
Uses medical entity phrases as carriers for multi-modal prompting and replaces unified cross-attention volume-phrase alignment with a dual-stream alignment that separates alignment paths (abstract is truncated, so the exact streams are not specified).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| split entity phrases into organ and lesion streams with separate cross-attention (ablation) | DEAR · CT-RATE | GREEN: 0.219 -> 0.247 | Improved |
| add organ-guided cross-attention (OGCA) to dual-stream aligner (ablation) | DEAR · CT-RATE | GREEN: 0.247 -> 0.439 | Improved |
| add lesion-guided cross-attention (LGCA) to dual-stream aligner (ablation) | DEAR · CT-RATE | GREEN: 0.247 -> 0.381 | Improved |
| combine OGCA and LGCA (ablation) | DEAR · CT-RATE | GREEN: 0.439 -> 0.511 | Improved |
| add Top-k voxel token selection in OGCA (ablation) | DEAR OGCA · BIMCV-R | GREEN: 0.287 -> 0.366 | Improved |
| use 3D dilation factor 3 instead of 1 (ablation) | DEAR OGCA · BIMCV-R | GREEN: 0.312 -> 0.366 | Improved |
| use 3D dilation factor 3 instead of 4 (ablation) | DEAR OGCA · BIMCV-R | GREEN: 0.348 -> 0.366 | Improved |
| add residual 3D Laplacian filtering in LGCA (ablation) | DEAR LGCA · CTRG-Chest | GREEN: 0.448 -> 0.463 | Improved |
| use Laplacian filtering without residual connection (ablation) | DEAR LGCA · CTRG-Chest | Precision: 0.463 -> 0.452 | Hurt |
| use DEAR instead of fVLM (unified alignment) | 3D radiology report generation · CT-RATE | GREEN: 0.401 -> 0.511 | Improved |
| use DEAR instead of CT2-Rep | 3D radiology report generation · CT-RATE | BLEU-4: 0.369 -> 0.366 | No clear effect |
Closest papers by meaning
- MEPNet: Medical Entity-Balanced Prompting Network for Brain CT Report Generation (AAAI 2025 · similarity 0.88)
- S2D-Align: Shallow-to-Deep Auxiliary Learning for Anatomically-Grounded Radiology Report Generation (AAAI 2026 · similarity 0.87)
- Anatomical Region-Guided Contrastive Decoding: A Plug-and-Play Strategy for Mitigating Hallucinations in Medical VLMs (AAAI 2026 · similarity 0.86)
- DAMPER: A Dual-Stage Medical Report Generation Framework with Coarse-Grained MeSH Alignment and Fine-Grained Hypergraph Matching (AAAI 2025 · similarity 0.86)
- CURE: Curriculum-guided Multi-task Training for Reliable Anatomy Grounded Report Generation (CVPR 2026 · similarity 0.85)
- MMedPO: Aligning Medical Vision-Language Models with Clinical-Aware Multimodal Preference Optimization (ICML 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.
