Causal Modelling with Graph Neural Networks for Personalised Medicine in Computational Pathology

“这个病人患的是哪种类型的乳腺癌？”，“这个病人的肿瘤中有哪些突变？”，“化疗是否有助于提高这个病人的生存率？”计算病理学（CPath）通过使用人工智能（AI）分析数字扫描组织载玻片的数十亿像素整片图像（wsi），为回答这些问题提供了革命性的新方法。由于有望提供定量、客观和可重复的结果，计算病理学中的人工智能和机器学习（ML）方法将产生更有效的临床工作流程，并有助于克服当前和未来的挑战，这些挑战是由几乎所有发达国家的患者数量不断增加的工作量和病理学家队伍规模的减少所带来的。CPath可以协助病理学家、肿瘤学家和制药行业进行各种诊断和预后任务，以及为癌症患者选择和开发有效的个性化治疗方法。CPath的这些令人兴奋的可能性也带来了许多科学和计算方面的挑战。其中包括减少训练“数据匮乏”的人工智能方法对大量专家注释数据的需求，提高人工智能方法对来自不同中心和人群的数据变化的鲁棒性，使人工智能能够模拟组织图像的多分辨率性质，以捕获与诊断、预后和疾病结果相关的有意义的组织学特征，并确保人工智能方法提供可解释和可操作的结果。虽然CPath目前是一个非常活跃的研究领域，但该领域的大多数现有方法都无法以最小的训练数据需求以整体方式对wsi进行建模。此外，该领域的现有方法没有明确地对工作中的潜在因果机制进行建模，以提供反事实的解释（例如，“如果组织切片染色不同，这个机器学习模型的输出将如何变化？”）。或回答具有临床意义的反事实问题（例如，“如果这个病人接受不同的治疗会发生什么？”）。在这个项目中，我们将以工具箱的形式开发计算方法，这将有助于克服这些缺点，并为精准医疗产生有效的人工智能。具体来说，研究团队将开发基于图神经网络（gnn）的方法，该方法可以通过学习wsi的有效表示来模拟大型整个幻灯片图像的细胞拓扑和空间异质性，而不需要大量的训练数据。这些gnn将与因果模型相结合，以提供反事实解释，并提高人工智能方法对源于与潜在疾病或治疗机制不直接相关的因素的非因果变化的鲁棒性。在这项研究中开发的人工智能工具将为个性化医疗的临床重要问题提供有效的解决方案。特别是，该研究将能够从常规组织学图像中预测乳腺癌患者的受体状态，这将减少与治疗选择这一基本临床步骤相关的等待时间和成本。此外，它将使人们更深入地了解患者肿瘤组织图像中的哪些因素可以预测他们对治疗的反应。因此，拟议的研究将产生新颖有效的CPath技术，并在这一新兴领域开辟一条以前未探索的因果建模途径。根据EPSRC的使命，拟议的研究将通过技术发展和培训高技能劳动力，帮助确保英国在医疗保健和计算病理学商业上可行的人工智能领域的领导能力。这也符合2019年NHS长期计划中改善人工智能和数字医疗技术使用的国家战略优先事项

"What type of breast cancer does this patient have?", "What are the mutations in the tumour of this patient?", "Will chemotherapy help improve survival for this patient?" - Computational pathology (CPath) is providing revolutionary new ways of answering such questions by using Artificial Intelligence (AI) for analysis of multi-gigapixel whole slide images (WSIs) of digitally scanned tissue slides. With the promise of providing quantitative, objective and reproducible results, AI and machine learning (ML) approaches in computational pathology will yield more efficient clinical workflows and help overcome current and future challenges posed by an ever-increasing workload in terms of number of patients and decrease in the size of the pathologist workforce in almost all developed countries. CPath can assist pathologists, oncologists and the pharmaceutical industry in various diagnostic and prognostic tasks as well as the selection and development of effective personalized treatments for cancer patients. These exciting possibilities of CPath also come with many scientific and computational challenges. These include reducing the requirement of large amounts of expertly-annotated data for training "data-hungry" AI methods, improving robustness of AI approaches to variations in data from different centres and populations, enabling AI to model the multiresolution nature of tissue images to capture meaningful histological characteristics associated with diagnosis, prognosis and disease outcomes, and ensuring that AI methods provide explainable and actionable results. While CPath is currently a very active research field, most existing approaches in this domain are unable to model WSIs in a holistic manner with minimal training data requirements. Furthermore, no existing approaches in this domain explicitly model the underlying causal mechanisms at work to provide counterfactual explanations (e.g., "How will the output of this machine learning model change if the tissue slide was stained differently?") or answer counterfactual questions of clinical significance (e.g., "What would have happened had this patient been given a different treatment?"). In this project, we will develop computational approaches in the form of toolboxes that will help overcome these shortcomings and produce effective AI for precision medicine. Specifically, the research team will develop methods based on graph neural networks (GNNs) which can model cellular topology and spatial heterogeneity in large whole slide images by learning effective representations of WSIs without requiring large amounts of training data. These GNNs will be integrated with causal modelling to provide counterfactual explanations and improve robustness of AI methods to non-causal variations stemming from factors that are not directly related to underlying disease or treatment mechanisms. The AI tools developed in this research will deliver effective solutions to clinically important problems in personalised medicine. In particular, the research will enable prediction of receptor status of breast cancer patients from routine histology images which will reduce waiting times and costs associated with this fundamental clinical step in treatment selection. Furthermore, it will enable a deeper understanding of what factors in the tissue image of a patient's tumour are predictive of their response to treatment. The proposed research will thus result in novel and effective CPath technologies and open up a previously unexplored avenue of causal modelling in this emerging field. In line with EPSRC's mission, the proposed research will help ensure the UK's leadership capacity in the field of AI in healthcare and the commercially viable area of computational pathology through technological development as well as training of a highly skilled workforce. It also aligns with the national strategic prioritization of improved use of AI and digital healthcare technologies in the 2019 NHS Long Term Plan