SMR-R1 Dataset
收藏资源简介:
我们开源了一个医疗报告结构化数据评估集。该数据集具有以下特点:多样性和代表性:数据集包含各种类型的医疗报告,如病历、检查报告和诊断报告。它涵盖了一系列影像条件,包括透视、斜视图和光照强度的变化。此外,还包括非医疗报告相关的图像(如药盒)以评估模型过滤非医疗图像的能力。高质量标注:所有数据均由专业医疗人员标注和审核,确保准确性和可靠性。数据匿名化:为保护患者隐私,数据集中的所有信息均已匿名化,确保不包含个人身份信息。
We have open-sourced a structured medical report evaluation dataset. This dataset features the following attributes: 1. Diversity and representativeness: The dataset encompasses various types of medical reports, including medical records, examination reports and diagnostic reports. It covers a wide range of imaging conditions, such as perspective views, oblique views and variations in lighting intensity. Additionally, it includes non-medical report-related images (e.g., medicine boxes) to evaluate the model's ability to filter out non-medical images. 2. High-quality annotations: All data are annotated and reviewed by professional medical personnel to ensure accuracy and reliability. 3. Data anonymization: To protect patient privacy, all information within the dataset has been anonymized, ensuring that no personal identifiable information (PII) is present.
SMR-R1数据集概述
数据集简介
- 名称: SMR-R1
- 目的: 提供医疗报告结构化信息提取的解决方案
- 内容:
- 脱敏医疗报告结构化数据评估集
- 医疗报告结构化提取模型
- 医疗报告结构化提取评估方法
数据集特点
- 多样性: 包含多种医疗报告类型(病历、检查报告、诊断报告等)
- 代表性: 涵盖多种影像条件(透视、斜位视图、光照强度变化)
- 数据质量: 由专业医疗人员标注和审核
- 隐私保护: 所有数据经过匿名化处理
模型信息
- 基础模型: Qwen2.5VL-7B
- 训练方法: GRPO算法强化训练
- 性能提升:
- 比原始Qwen2.5VL-7B模型提高15个百分点
- 比相同数据SFT训练高15个百分点
- 比Qwen2.5VL-72B模型高7个百分点
评估结果
| 模型 | 数据集 | 性能 | 说明 |
|---|---|---|---|
| Qwen2.5-VL-7B-Instruct | - | 54.71% | - |
| Qwen2.5-VL-72B-Instruct | - | 62.89% | - |
| Qwen2.5-VL-7B-Instruct-sft | 6k | 55.21% | 监督微调 |
| SMR-R1-7B | 6k | 71.87% | 强化学习 |
数据示例
json { "姓名": "", "性别": "", "年龄": "", "检查时间": "", "科室": "", "检查名称": "", "检查部位": "", "指标": [ { "项目名称": "免疫球蛋白G [IgG]", "结果": "12.10", "单位": "g/L", "参考范围": "5.28--21.9", "异常标记": "", "检测方法": "", "结果状态": "正常" } ] }
未来计划
- [ ] 在基础模型SFT后进行强化学习
- [ ] 不限制提取字段,实现医疗报告自动结构化
- [ ] 优化奖励方法(将由大型模型确定)
引用信息
bib @misc{lijun2025SMR-R1, author = {Lijun Liu, Tao Zhang, Tao Zhang, Chong Li, Mingrui Wang, Chenglin Zhu, Mingan Lin, Zenan Zhou, Weipeng Chen}, title = {SMR-R1: Reinforcing Ability to Extract Structured Information From Medical Reports in Vision Language Models}, howpublished = {url{https://github.com/yingyukexiansheng/SMR-R1}}, note = {Accessed: 2025-03-26}, year = {2025} }




