AI-based diagnosis for improving classification of bone and soft tissue tumours across the UK

病理组织诊断的交付，其中大多数是癌症，在目前的格式是不可持续的。基因组医学和免疫肿瘤学的进展表明，将肿瘤分类为亚型可以选择特定治疗的患者，但也可以使患者免于不必要的毒性和昂贵的治疗。尽管如此，做出这样的诊断已经变得更加耗时，包括选择和解释辅助测试，这需要对每种癌症类型不断增长的专业知识。虽然对诊断专家的需求正在增加，但能够报告结果的病理学家已经短缺了25%：这一比例将会下降。我们建议使用人工智能可以确保病理学家提供的组织诊断是可持续的，并支持个性化治疗的提供。人工智能在病理学方面的好处已经开始显现，例如，前列腺癌高级别区域的识别可以减少错误和病理学家的时间。人工智能在诊断方面的发展是及时的，因为预计到2025年，英国将全面采用数字化组织图像，使它们能够同时由人类和人工智能（AI）进行查询。人工智能是一个需要大量数据的过程；提供100,000张图像来训练模型是不现实的。即使是最常见的癌症（如乳腺癌）也有多种亚型；识别这些是选择患者进行个性化治疗的必要条件。为了应对这一挑战，我们建议使用相对较小的样本量（每个类约1000张图像）开发一种新的AI策略。这样的模型可以适用于任何类型的癌症。将开发一个多实例学习框架，使用变压器进行特征提取和分类。将应用一种标记无法确定分类的样本的工具，从而提醒病理学家注意潜在的未见疾病。深度学习模型将通过注入病理学家的领域知识得到加强。软组织和骨骼肿瘤我们将开发软组织（肌肉、脂肪、血管等）和骨骼肿瘤的人工智能模型，这被认为是最具挑战性的诊断领域之一。这些肿瘤包括大约100种不同的亚型，代表了儿童和年轻人中最常见的一些癌症。我们将建立现有的深度学习模型，该模型在2122张图像上训练了15种不同的亚型，该模型预测了87%的病例的正确诊断。然后由算法提示确定辅助测试的选择，并简化诊断途径。已经扫描的17000张图像将被添加到库中，并允许分类模型的快速发展和扩展。该图片库将与临床结果相关联，并在项目期间扩展到35000张图像。除此之外，由来自英国所有国家的至少20名病理学家组成的肉瘤网络承诺提供上述额外的20,000张图像。该研究和基础设施将作为该模型持续发展的框架，随着数字病理学在NHS和全球的引入，该模型可以迅速扩展。该模型可以随着时间的推移而发展，以响应新的进展。该图像库将用于培训未来的病理学家、研究、验证其他人工智能算法，并为肉瘤基因组学英国临床解释合作伙伴关系（GeCIP）提供宝贵的资源，为未来的多模态多组学研究做出贡献。我们将与肉瘤慈善机构和合作伙伴密切合作，让患者、他们的家人和公众参与进来，建立对人工智能在医疗保健领域使用的信任。开发用于数字化病理图像的人工智能模型可以避免这一医学专业面临的危机。

Delivery of pathology tissue diagnoses, most of which are cancer, in the current format is unsustainable. Advances in genomic medicine and immune-oncology have shown that the classification of tumours into subtypes allows selection of patients for specific treatments but also spares patients unnecessary toxic and expensive therapies. Still, making such diagnoses has become more time-consuming, involving the selection and interpretation of ancillary tests which requires an ever-growing specialist knowledge for each cancer type. Whilst the need for diagnostic expertise is increasing, there is already a shortfall of 25% of pathologists who are able to report results: this is set to decline.We propose that the use of AI can ensure that the delivery of tissue diagnoses by pathologists is sustainable and supports delivery of personalised treatments. The benefits of AI in pathology are beginning to be seen, e.g. identification of high-grade areas of prostate cancer shows a reduction in errors and pathologists' time. The development of AI for diagnoses is timely as full adoption of digitised histological images, allowing them to be interrogated by both humans and artificial intelligence (AI), is expected in the UK by 2025. AI is a data-hungry process; it is unrealistic to provide 100,000s images that are required to train a model. Even the most common cancers (e.g. breast) have multiple subtypes; identification of these is required for selection of patients for personalised treatments. To address this challenge, we propose to develop a novel AI strategy using a relatively small sample size (~1000 images per class). Such a model could be adapted to any cancer type. A multiple-instance learning framework will be developed, using transformers for feature extraction and classification. A tool that flags samples that cannot be confidently classified will be applied thereby alerting the pathologist of potentially unseen diseases. The deep learning model will be strengthened by the injection of pathologists' domain knowledge. Soft tissue and bone tumoursWe will develop the AI model on tumours of soft tissue (muscle, fat, blood vessels, etc.) and bone, an area considered to be one of the most challenging diagnostically. These tumours comprise approximately 100 different subtypes, and represent some of the most common cancers in children and young adults. We will build on our existing deep learning model of 15 different subtypes trained on 2122 images, which predicts the correct diagnosis in 87% of cases. Selection of confirmatory ancillary tests is then prompted by the algorithm and streamlines the diagnostic pathway. 17,000 images that have already been scanned will be added to the library and allow the rapid development and extension of the classification model. The image library will be linked to clinical outcomes and expanded to 35,000 images during the project. Added to this is the commitment of the established Sarcoma Network of at least 20 pathologists from across all countries in the UK, to provide the additional 20,000 images mentioned above. Additional benefitsThe study and infrastructure will serve as the framework for the continued development of the model which can rapidly be expanded prospectively with the introduction of digital pathology in the NHS and globally. The model can be developed over time in response to new advances. The image library will be available for training future pathologists, research, validation of other AI algorithms, and contribute to the Sarcoma Genomics England Clinical Interpretation Partnership (GeCIP) offering a valuable resource for future multi-modal multi-omic research.Working closely with Sarcoma charities, and partners, we will involve and engage patients, their families, and the public, to build trust in the use of AI in health care. Development of AI models for digitised pathology images can avert the crisis facing this medical specialty.