Research map / Vision-language and multimodal / CLIP and zero-shot
LLM-Guided Diagnostic Evidence Alignment for Medical Vision–Language Pretraining under Limited Pairing
Huimin Yan, Liang Bai, Xian Yang, Long Chen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use LGDEA with only 10% paired data instead of fully paired local-alignment baselines | LGDEA · MS-CXR phrase grounding (8 categories) | CNR: Atelectasis 0.8312 (MedAligner) -> 0.8449 | Improved |
| increase paired data from 5% to 10% (ablation) | LGDEA · MS-CXR phrase grounding, Pleural Effusion | CNR: 0.9742 -> 1.0928 | Improved |
| reduce paired data for baselines to 50% (ablation) | MedAligner, GLoRIA, MGCA · MS-CXR phrase grounding | CNR: MedAligner Atelectasis 0.8312 -> 0.6346 | Hurt |
| add unpaired CheXpert images (cross-domain pretraining) (ablation) | LGDEA 10% pair · MS-CXR phrase grounding, Pleural Effusion | CNR: 1.0928 -> 0.7519 | Hurt |
| use LGDEA with 10% paired data vs MedAligner with full pairs | LGDEA · MIMIC-5x200 retrieval | Prec@2: 54.08 -> 53.85 | No clear effect |
| use LGDEA with 10% paired data vs MedAligner with full pairs | LGDEA · RSNA Pneumonia classification | accuracy (%): 80.05 -> 79.26 | Hurt |
| add evidence prototype learning L_rec (ablation) | LGDEA · downstream benchmarks | performance | Improved |
| replace Spark-Desk with Qwen-7B or LLaMA-8B for evidence extraction (ablation) | LGDEA · zero-shot classification | accuracy (%): NIH 87.06 -> 86.14 (Qwen), 85.80 (LLaMA) | No clear effect |
| increase diagnostic prototypes K from 32 to 64 (ablation) | LGDEA · COVID classification | accuracy (%): 87.14 -> 90.47 | Improved |
| increase diagnostic prototypes K beyond 64 (ablation) | LGDEA · COVID classification | accuracy (%): 90.47 -> 84.72 (K=128) | Hurt |
| increase lesion queries L from 32 to 64 (ablation) | LGDEA · COVID classification | accuracy (%): 85.10 -> 90.47 | Improved |
| increase lesion queries L beyond 64 (ablation) | LGDEA · COVID classification | accuracy (%): 90.47 -> 85.57 (L=80) | Hurt |
Closest papers by meaning
- LLM-CXR: Instruction-Finetuned LLM for CXR Image Understanding and Generation (ICLR 2024 · similarity 0.88)
- Enhancing Medical Large Vision-Language Models via Alignment Distillation (AAAI 2026 · similarity 0.88)
- HC-LLM: Historical-Constrained Large Language Models for Radiology Report Generation (AAAI 2025 · similarity 0.88)
- ExGra-Med: Extended Context Graph Alignment for Medical Vision-Language Models (NeurIPS 2025 · similarity 0.87)
- MedSG-Bench: A Benchmark for Medical Image Sequences Grounding (NeurIPS 2025 · similarity 0.87)
- S2D-Align: Shallow-to-Deep Auxiliary Learning for Anatomically-Grounded Radiology Report Generation (AAAI 2026 · similarity 0.87)
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.
