A DeepSeek-Powered AI System for Automated Chest Radiograph Interpretation in Clinical Practice
试验速览
- 阶段
- 不适用
- 状态
- 已完成
- 发起方
- 入组人数
- 296
- 试验地点
- 4
- 主要终点
- Report quality scores in the prospective study
研究概览
简要总结
There's a global shortage of radiologists. Radiology AI's automatic reporting is key for boosting efficiency and meeting patient needs, especially in resource-poor areas. Multimodal large models enable medical image auto-reporting systems. ChatGPT 4o can diagnose medical images but has issues like being closed-source and "hallucinations." The new open-source Janus Pro 1B-with strong performance, "any-to-any" capability, low cost, and open access-shows potential for medical imaging tasks with training. But little research explores its use here; most models are general, lacking field-specific optimization and systematic evaluation. This study will develop Janus Pro 1B-CXR (a medical image-specific model) via public data, test its value in diagnosis and reporting, and build an efficient automated system.
详细描述
There is a global shortage of radiologists, and the automatic report generation function of radiology AI systems is crucial for improving medical efficiency and meeting patient needs, especially those in areas with scarce medical resources. Multimodal large models have made it possible to develop automatic report generation systems for medical images. Although ChatGPT 4o has certain capabilities in medical image diagnosis, it has issues such as being closed-source and hallucination. The recently launched open-source multimodal large model Janus-Pro has advantages including high performance, "Any to any", low cost, and open-source; after training and fine-tuning, it has the potential for medical image diagnosis and report generation. However, there is currently a lack of research on the application of Janus Pro 1B in image diagnosis; existing models are mostly general-purpose, lacking in-depth optimization for specific fields and systematic multi-dimensional evaluation methods. This study aims to develop a large model specialized in medical images, Janus Pro 1B-CXR, using public databases, verify its application value in image diagnosis and radiology report generation, and construct an efficient and accurate automated medical image analysis and diagnostic assistance system.
研究设计
- 研究类型
- Interventional
- 分配方式
- Randomized
- 干预模型
- Parallel
- 主要目的
- Diagnostic
- 盲法
- Double (Participant, Outcomes Assessor)
入排标准
- 性别
- All
- 接受健康志愿者
- 是
入选标准
- •clinically suspected thoracic diseases (such as pneumonia, tuberculosis, or lung cancer) requiring CXR-assisted diagnosis;
- •patients providing written informed consent for research data use;
- •complete clinical records (including chief complaints, medical history, and laboratory test results);
- •patients with no historical chest X-ray images and no need for comparison with previous chest X-ray images;
- •Patients who underwent only posteroanterior (PA) chest X-rays without lateral chest X-rays.
排除标准
- •substandard CXR image quality (including severe motion artifacts, over-/underexposure, or missing anatomical structures);
- •pregnant or lactating women.
结局指标
主要结局
Report quality scores in the prospective study
时间窗: 1 week
In this prospective study, the quality of reports generated by junior radiologists was assessed using a 5-point Likert scale titled "Radiology Report Quality Assessment Scale", where the minimum value is 1 and the maximum value is 5, with higher scores indicating better report quality. These scores were compared between the AI-assisted group (junior radiologists using AI tools for report generation) and the standard care group (junior radiologists generating reports without AI assistance).
Agreement evaluation in the prospective study
时间窗: 1 week
In this prospective study, the agreement between reports generated by junior radiologists and standard reports was assessed using the RADPEER scoring system-a peer review program established by the American College of Radiology (ACR) designed to evaluate the interpretation accuracy of radiologists-where the degree of concordance is measured by grading discrepancies and agreements according to specific criteria that also account for the clinical significance of differences. The RADPEER system uses a 5-category scale with a minimum value of 1 and a maximum value of 5, where higher scores indicate greater agreement between the generated reports and standard reports.
Pairwise preference tests in the prospective study
时间窗: 1 week
In this prospective study, the preference between reports generated by junior radiologists in the AI-assisted group versus the standard care group was evaluated using the "Expert Pairwise Preference Assessment Tool", a structured measurement tool designed to quantify expert consensus on report superiority. The assessment was conducted by a panel of 5 independent radiology experts, who reviewed paired reports (one from the AI-assisted group and one from the standard-care group for the same clinical case) and individually indicated their preference for which report was more clinically valuable, accurate, or comprehensive. The unit of measure for this outcome is the "Percentage of paired cases with majority expert preference", defined as cases where ≥3 out of 5 experts expressed a clear preference for either the AI-assisted or standard care report.
Reading Time in the prospective study
时间窗: 1 week
The time from when radiologists began examining chest radiographs to the completion of final reports, comparing efficiency between the AI-assisted and Standard-care groups.
次要结局
- Report Quality Score in the retrospective study(1 week)
- Agreement Evaluation in the retrospective study(1 week)
- Pairwise preference tests in the retrospective study(1 week)
研究者
Yaowei Bai
Clinical Investigator
Union Hospital, Tongji Medical College, Huazhong University of Science and Technology