我们关于开放科学的立场与承诺
我们相信，科学研究的价值不仅在于发表论文，更在于推动知识积累、促进技术进步，并最终惠及患者。对于口腔医学AI研究而言，开放科学不仅有助于提高研究的可重复性和透明度，也能够降低技术门槛，让更多研究人员和临床医生从人工智能的发展中受益。
我们支持开放、可重复和可持续的科研实践，同时认为，负责任的开放科学应当兼顾科学共同体的公共利益与原始研究者的合理权益。基于这一认识，我们提出以下关于口腔AI研究成果开放的观点与承诺。
1. 我们支持尽可能充分的研究成果开放
我们认为，开源不应仅仅停留在论文中的一句声明，而应当使研究成果真正能够被理解、验证、使用和改进。
对于条件允许的研究，我们鼓励公开完整的源代码、预训练模型权重、经过适当处理的原始数据或高质量合成数据，以及必要的技术文档和实验配置。
尤其是预训练模型权重，其开放能够显著降低其他研究人员复现与拓展研究的技术和计算成本。对于依赖大量专家标注和模型训练的医学AI研究，这一点尤为重要。
我们也鼓励提供便于临床医生和非计算机专业研究人员使用的在线演示系统（Demo），使其能够在无需配置复杂开发环境的情况下体验、测试和评价模型的实际表现。
我们的目标不仅是让研究成果可以被获取，更是让它们真正能够被使用。
2. 对于承诺开源的代码，我们主张在论文在线出版后立即公开
科学论文的发表意味着研究成果正式进入公共知识体系。对于已经承诺开源代码的研究，我们认为，代码应当与论文尽可能同步公开，而不应长期停留在“Code will be available soon”的状态。
我们主张并承诺，对于明确声明将公开源代码的研究项目，原则上在论文首次在线出版后的3天内完成代码开放。
公开的代码应尽可能包括模型实现、必要的训练与推理脚本、依赖环境说明及使用文档，以便其他研究人员理解和复现所提出的方法。
我们认为，一项明确而可执行的开源承诺，比没有时间表的模糊承诺更有价值。
3. 对于数据和预训练模型权重，我们认为有限期限的延迟开放是合理的
高质量医学AI研究往往建立在大量研究劳动之上。患者数据采集、临床专家标注、数据质量控制、算法开发及模型训练，都可能需要研究团队投入数年时间和大量资源。这些投入不仅支撑了一篇论文，也常常为后续一系列研究奠定基础。
因此，我们认为，开放科学不应忽视原始研究团队的合理权益，特别是那些在数据建设、专家标注和基础模型开发中作出实质贡献的研究人员。他们应当获得合理的时间窗口，优先开展基于这些成果的后续研究。
这一观点并非没有先例。
2015年，美国国家科学院出版社出版的报告 Sharing Clinical Trial Data: Maximizing Benefits, Minimizing Risk 明确建议，支持已发表临床试验结果的分析数据应不晚于论文发表后6个月共享。报告同时指出，这一期限有助于平衡快速共享数据的公共利益与原始研究者撰写后续论文时应有的竞争优势。
我们认为该建议背后的科研治理原则同样值得医学AI领域借鉴，开放科学应当鼓励知识共享，也应当尊重知识生产者的劳动和合理权益。
基于这一原则，对于明确承诺公开数据（包括合成数据）或预训练模型权重的项目，我们主张在论文首次在线出版后不超过6个月的期限内完成公开。对于采用延迟开放政策的项目，我们将在论文或项目页面中明确说明并承诺开放计划、预计日期及访问方式，避免无限期延迟。
我们希望通过这种有限、透明且可预期的安排，在促进开放科学的同时，为原始研究团队及其合作者保留合理的后续研究机会。
4. 我们希望建立一种更加公平的开放科学文化
我们认为，科研成果的共享与原始研究者的权益并不矛盾。
开放科学需要更多人愿意投入资源建设高质量数据集、开发基础模型和维护公共研究工具。因此，学术共同体不仅应当鼓励使用开放资源，也应当充分认可资源生产者、数据标注者、临床合作者及算法开发者的贡献。
我们支持对开放数据、模型和代码给予规范引用与适当学术认可，也支持研究团队通过透明、合理的内部机制，为参与基础资源建设的研究人员提供优先开展后续研究的机会。
我们相信，真正可持续的开放科学文化，应当让开放成果的研究者获得应有的认可，也让使用成果的人能够不断创造新的价值。
5. 结语
我们支持开放科学，也尊重科学研究背后的劳动。我们的基本立场是：尽可能开放，明确开放承诺，及时兑现承诺，并在开放共享与研究者合理权益之间保持公平的平衡。对于承诺开源的项目，除非特别说明，我们将分别遵循以下原则：
- 源代码：论文首次在线出版后3天内公开。
- 数据及预训练模型权重：论文首次在线出版后6个月内公开。
- 在线Demo：在技术与伦理条件允许时，尽可能向研究和临床社区提供便捷的测试途径。
我们希望，口腔AI研究能够发展出一种更加开放、透明、公平且可持续的科研文化，让每一项研究投入都能够创造超越单篇论文的长期价值。
We believe that the value of scientific research lies not only in publishing papers, but also in advancing knowledge, enabling technological progress, and ultimately benefiting patients. In oral healthcare AI research, open science can improve reproducibility and transparency, lower technical barriers, and help more researchers and clinicians benefit from advances in artificial intelligence.
We support open, reproducible, and sustainable research practices. We also believe that responsible open science should balance the public interest of the scientific community with the legitimate interests of the researchers who produced the original work. On this basis, we set out the following positions and commitments on sharing research outputs in oral healthcare AI.
1. We support sharing research outputs as fully as possible
We believe that a commitment to open source should go beyond a statement in a paper: research outputs should be genuinely understandable, verifiable, usable, and open to improvement.
Where conditions permit, we encourage the release of complete source code, pretrained model weights, appropriately processed original data or high-quality synthetic data, and the technical documentation and experimental configurations needed to use them.
Sharing pretrained model weights is particularly valuable because it can substantially reduce the technical and computational costs of reproducing and extending research. This is especially important in medical AI research that depends on extensive expert annotation and model training.
We also encourage online demonstration systems (demos) that are accessible to clinicians and researchers without a computing background, allowing them to try, test, and evaluate models without setting up a complex development environment.
Our goal is to make research outputs not only available, but genuinely usable.
2. When source code has been promised, we advocate releasing it immediately after online publication
Publication formally brings research findings into the public body of scientific knowledge. For studies that have committed to releasing their code, we believe that the code should be made available as close to publication as possible, rather than remaining indefinitely at the stage of “Code will be available soon.”
For research projects that explicitly commit to releasing source code, we advocate and commit, in principle, to making it public within three days of the paper’s first online publication.
Where possible, the release should include the model implementation, the necessary training and inference scripts, environment and dependency information, and usage documentation, so that other researchers can understand and reproduce the proposed methods.
We believe that a clear, actionable commitment to open source is more valuable than a vague promise without a timeline.
3. We consider a limited delay in releasing data and pretrained model weights reasonable
High-quality medical AI research often rests on substantial research effort. Collecting patient data, obtaining clinical expert annotations, ensuring data quality, developing algorithms, and training models can require years of work and considerable resources. These investments often support not only a single paper, but also a series of subsequent studies.
We therefore believe that open science should take account of the legitimate interests of the original research team, particularly those who have made substantial contributions to data collection and curation, expert annotation, and foundation model development. They should have a reasonable window of time to pursue follow-up research using these resources.
There is precedent for this position.
The 2015 report Sharing Clinical Trial Data: Maximizing Benefits, Minimizing Risk, published by the National Academies Press, recommended that the analytical data supporting published clinical trial results be shared no later than six months after publication. The report also noted that this timeframe helps balance the public benefit of rapid data sharing with the competitive advantage that original investigators should have when preparing subsequent publications.
We believe that the research governance principle underlying this recommendation is also relevant to medical AI: open science should encourage knowledge sharing while respecting the work and legitimate interests of those who produce that knowledge.
On this basis, for projects that explicitly commit to releasing data, including synthetic data, or pretrained model weights, we advocate making these resources public no later than six months after the paper’s first online publication. When a project adopts a delayed release policy, we will clearly state the release plan, expected date, and access arrangements in the paper or on the project page, and commit to that plan, avoiding indefinite delays.
Through this limited, transparent, and predictable arrangement, we hope to advance open science while preserving reasonable opportunities for the original research team and its collaborators to pursue follow-up studies.
4. We hope to build a fairer culture of open science
We believe that sharing research outputs and respecting the interests of the original researchers are compatible.
Open science depends on people being willing to invest resources in building high-quality datasets, developing foundation models, and maintaining public research tools. The academic community should therefore encourage the use of open resources while fully recognizing the contributions of resource creators, data annotators, clinical collaborators, and algorithm developers.
We support proper citation and appropriate academic recognition for open data, models, and code. We also support transparent and reasonable arrangements within research teams that give contributors to foundational resources priority opportunities to conduct follow-up research.
We believe that a sustainable culture of open science should give due recognition to researchers who share their work and enable those who use it to continue creating new value.
5. Closing statement
We support open science and respect the work behind scientific research. Our fundamental position is to share as openly as possible, make clear commitments, deliver on them promptly, and maintain a fair balance between open sharing and the legitimate interests of researchers. Unless otherwise stated, projects that commit to open release will follow these principles:
- Source code: released within three days of the paper’s first online publication.
- Data and pretrained model weights: released within six months of the paper’s first online publication.
- Online demos: where technical and ethical conditions permit, we will seek to provide convenient ways for the research and clinical communities to test our models.
We hope that oral healthcare AI research will develop a more open, transparent, fair, and sustainable research culture, allowing every investment in research to create lasting value beyond a single paper.