Artificial intelligence-driven radiosurgery for brain metastases

该项目的目标：实施基于学习的注册，以管理以前用于向脑转移患者提供立体定向放射手术的MR图像的空间归一化数据集；开发数据驱动的深度学习框架，以自动检测和分割脑转移，同时允许交互式纠正；开发成像生物标志物，以预测治疗后的肿瘤反应和行为。高达40%的颅外原发癌患者发生脑转移。此外，尽管目前没有可靠的数据，但转移性脑肿瘤的数量被认为至少是原发性恶性脑肿瘤的3:1。脑转移患者需要个性化的患者管理，可能包括手术、立体定向放射外科手术、分次放疗和化疗，无论是单独还是联合。脑转移最常发生在肺癌、乳腺癌、肾癌、黑色素瘤或肠癌患者身上。手术、立体定向放射外科（SRS）和全脑放疗（WBRT）仍然是脑转移治疗的主要方法。特别是，SRS已成为治疗颅内转移瘤的重要方式。选择SRS或WBRT的最重要标准是转移的总数量。SRS是1-10例转移患者的首选治疗方式，尽管治疗多发转移的临床医生数量正在增加，一些中心常规治疗超过10-20例转移。SRS治疗的计划受到转移数量的显著影响。由于必须仔细检查成像数据以检测和分割所有肿瘤，小的微转移的存在可以大大增加计划时间的长度。此外，所有脑转移患者都需要终身放射随访，即使在接受治疗后也是如此，这给医疗资源带来了额外的负担。自动分割工具可以显著改善SRS计划过程中的临床工作流程。通过在临床医生驱动的交互过程中使用人工智能分割工具作为初始化步骤，我们将改善工作流程和操作效率。人工智能（AI）是指模拟与人类智能相关的过程的计算技术。我们之前开发了一个全自动的人工智能框架，用于从MRI中分割前庭神经鞘瘤（另一种脑肿瘤），获得了最先进的结果。该项目目前旨在开发深度学习模型：1)使用MRI检测并自动分割脑转移瘤，同时允许用户驱动的纠正；2)开发复合临床和成像生物标志物，以预测伽玛刀治疗后肿瘤的反应和行为。该项目的数据集和学习将为一项雄心勃勃的研究计划提供基础，该计划旨在通过使人工智能工具足够灵活以无缝集成到临床工作流程中，将这些工具应用于临床实践。这将需要设计交互式更正，提供可解释性手段，以及开发经过验证的人工智能可信度特征。

Aims of the Project:Implement learning-based registration to curate a spatially-normalised dataset of MR images previously used to deliver stereotactic radiosurgery to brain metastasesDevelop data-driven deep learning frameworks to automatically detect and segment brain metastases while allowing for interactive correctionsDevelop imaging biomarkers to predict tumour response and behaviour following treatmentApproximately 25,000 patients are diagnosed with a brain tumour every year in the UK. Brain metastases affect up to 40% of patients with extracranial primary cancer. Furthermore, although there are presently no reliable data, metastatic brain tumours are thought to outnumber primary malignant brain tumours by at least 3:1. Patients with brain metastases require individualized patient management and may include surgery, stereotactic radiosurgery, fractionated radiotherapy and chemotherapy, either alone or in combination.Brain metastases most commonly occur in patients with lung, breast, kidney, melanoma or bowel cancer. Surgery, stereotactic radiosurgery (SRS), and whole-brain radiotherapy (WBRT) continue to be the mainstay of treatment for brain metastasis. In particular, SRS has emerged as an important modality for treating intracranial metastases. The most important criteria for choosing SRS or WBRT is the overall number of metastases. SRS is the preferred treatment modality for patients with 1-10 metastases although the number of clinicians treating multiple metastases is increasing and several centres routinely treat over 10-20+ metastases. The planning of SRS treatment is significantly impacted by the number of metastases. The presence of small micro-metastases can greatly increase the length of planning time, as a result of having to carefully scrutinise the imaging data to detect and segment all the tumours. Furthermore, lifelong radiological follow-up is required for all patients with brain metastases, even after they undergo treatment, placing an additional burden on healthcare resources. An automated segmentation tool could significantly improve clinical workflow during the planning of SRS. By using an AI segmentation tool as an initialisation step in a clinician-driven interactive process, we will improve workflow and operational efficiency.Artificial intelligence (AI) refers to computing technologies that mimic processes associated with human intelligence. We have previously developed a fully-automated AI framework to segment a vestibular schwannoma (another type of brain tumour) from MRI achieving state-of-the-art results.This project now aims to develop deep learning models to: 1) detect and automatically segment brain metastases using MRI while allowing for user-driven corrections; and 2) develop composite clinical and imaging biomarkers to predict tumour response and behaviour following gamma knife treatment. The dataset and learning from this project will provide the foundation of an ambitious research programme aiming at translating such tools in clinical practice by making the AI tools flexible enough to seamlessly integrate into the clinical workflow. This will require the design of interactive corrections, the provision of interpretability means, and the development of proven AI trustworthiness features.