可信病理人工智能:从理论到实践

人工智能正在融入病理学研究的各个领域，但在临床实践中却遇到了诸多挑战，包括因医疗数据隐私保护而形成“数据孤岛”，不利于人工智能模型的训练；现有人工智能模型缺乏可解释性，导致使用者无法理解而难以形成人机互动；人工智能模型对多模态数据利用不足，致使其预测效能难以进一步提升等。为解决上述难题，我们建议在现有病理人工智能研究中引入可信人工智能技术:(1)数据安全共享，在坚持数据保护的基础上打破数据孤岛，使用联邦学习的方法、仅调用数据训练的结果而不上传数据本身，在不影响数据安全的情况下大大增加可用于训练的数据量；(2)赋予人工智能可解释性，使用图神经网络技术模拟病理医生学习病理诊断的过程，使得模型本身具有可解释性；(3)多模态信息融合，使用知识图谱技术对更为多样和全面的数据来源进行整合并深入挖掘分析，获得更准确的模型。相信通过此三方面的工作，可信病理人工智能技术可使病理人工智能达到可控可靠和明确责任，从而促进病理人工智能的发展和临床应用。

Trusted Artificial Intelligence for Pathology: From Theory to Practice

Artificial intelligence (AI) has gradually integrated into every aspect of pathology research. However, we also encounter some problems in the practical application of pathological artificial intelligence. 1. Research institutions attach importance to the protection of data privacy, which results in the emergence of data islands and is detrimental to our training of AI models. 2. The lack of interpretability of existing AI models leads to users' incomprehension and difficulty in human-computer interaction. 3. AI models make insufficient use of multi-modal data, making it difficult to further improve their predictive effectiveness. To address the above-mentioned challenges, we propose to introduce the latest technologies of trusted artificial intelligence (TAI) into existing research of pathological AI, which is embodied as the following: 1. Securely share data. We try to break data islands on the basis of adhering to data protection. We can use federated learning methods, only provide the results of data training without uploading the data itself, and greatly increase the amount of data that can be used for training without affecting the data security. 2. Give AI interpretability. The technology of graphic neural networks is used to simulate the process of pathologists' learning pathological diagnosis, making the model itself interpretable. 3. Fuse multimodal information. Use the technology of knowledge graph to integrate and deepen the analysis of more diverse and comprehensive data sources in order to derive more accurate models. Through the above three aspects, we can achieve reliable and controllable pathological AI and clear the responsibility through trusted pathological AI technology, so as to promote the development and clinical application of pathological AI.