Our ambition is to position AI to enable more widespread adoption of next generation technologies and to democratise access to precision medicine for all patients in the NHS. The strong foundations upon which that ambition is built is comprised of our multidisciplinary team, data access and organisation pipelines, and our evolving informatics infrastructure. Image Multidisciplinary manpower Hosting and developing a new 'digital-ready' workforce to support imaging departments of the future including data scientists, information experts, natural language processing experts, data systems engineers and transformational leads, in addition to typical clinical research personnel such as radiologists, research radiographers, physicists and trial managers. The hub is building a workforce to accelerate the discovery, development, testing and integration of AI and informatics into the NHS Data access and organisation Working closely with The Royal Marsden Data Stewards team and the Biomedical Research Centre Trusted Research Environment (BRIDgE) we will ensure seamless data flows within our robust and safe governance frameworks Working towards integrating tools to facilitate cohort selection, image anonymisation, image annotation and structured reporting, simultaneously supporting The Royal Marsden's new clinical Picture Archiving and Communication System (PACS) and Radiology Information System (RIS) Using infographics to improve communication with health care professionals and patients by building frameworks that allow the testing and implementation of new AI tools into imaging workflows to improve the patient experience, improve efficiency and enable integration of imaging biomarkers into routine patient care Informatics infrastructure The AI Hub will maintain and expand the capabilities of XNAT - an open-source imaging informatics software platform dedicated to imaging-based research. XNAT is the backbone data management system for our AI in imaging pipeline will also be developed to archive all our research data and to provide novel annotation tools to feed our radiomics and machine learning pipelines AI Hub Radiomics Capabilities Radiomics can provide unique insights into cancers by the systematic analysis of medical images to reveal imaging signatures that relate to disease and treatment outcomes Infrastructure The AI hub radiomics team focusses on data analytics and model building using cutting edge statistical and machine learning methods The team has access to high performance computing facilities to enable the use of computationally intensive modelling and validation methods, and GPU hardware for running deep-learning algorithms Deep connections with The Royal Marsden's medical imaging departments, and especially the Magnetic Resource (MR) research group, enable a joined-up approach to leverage the latest image acquisition developments End-to-end radiomics pipeline An end-to-end pipeline for developing and validating radiomics models has been created, built around an open-source python toolbox for generating radiomic features (pyradiomics) that is rapidly becoming the industry standard Manual tumour outlining is supported by web-based image visualisation tools within our XNAT platform, which also has automatic segmentation capabilities Algorithms for sub-segmentation of tumour images into biologically distinct regions are part of the pipeline, from which richer radiomic feature sets can be generated that capture a wide range of tumour heterogeneity The existing pipeline uses Python software for building classification models, R software for time-to-event (survival) studies and TensorFlow for deep learning studies, and almost any analysis software can be incorporated using docker containerisation Research topics of interest Developing radiomics biomarkers that provide solutions to areas of unmet need in the clinical decision pathway using scans and other information that would be routinely acquired Methodologies for development of radiomics models that are robust in real-world scenarios, e.g. to user-variation in tumour segmentation, scanning protocols, etc Strategies for generating interpretable radiomics models Linking radiomics to other -omics data; the generation of multi-omics signatures, and the use of other -omics data to assist in the interpretation of radiomic signatures Exemplar: Radiomics and deep learning algorithms for the assessment of post-chemotherapy residual mass in germ cell tumour For patients with testis germ cell tumours (GCT), one of the biggest issues is determining whether they will benefit from surgery after first-line chemotherapy. Patients frequently have residual masses that may contain active cancer, mature teratoma or necrosis/fibrosis. The first two need surgery as the curative intervention whereas those with necrosis will not benefit. To date there is no reliable way to distinguish between the two and so surgery is recommended for all patients with residual masses, despite the risks of surgical morbidity. Around 25-50% of tumours are found to contain only necrosis and thus patients are subjected to surgery with little additional benefit that comes with significant short and long terms harms, including a high risk of sexual dysfunction. Patients routinely have pre-surgical CT scans that are used in surgical planning, and experienced radiologists can often distinguish cases with residual disease from those without. However, such visual assessments can rarely be made with sufficient certainty to affect the decision to have surgery with curative intent. Radiomics analysis may offer a way to accurately characterise GCT residual masses noninvasively from pre-surgery CT scans, which would avoid a substantial proportion of these patients having complex major surgery with long term morbidity. A retrospective study is underway in the AI hub to explore this problem, and is a collaboration between the RMH urology surgical/radiology team, who have access to a list of eligible patient scans, and the AI hub. Manual segmentation is the most time-consuming step in this study, and the AI hub team includes two radiographers dedicated to this task. Radiographer training for this specific disease area has been provided by an experienced radiologist, who has also checked and finalised the segmentations. Work developing and validating a radiomics signature using features generated from the pyradiomics toolbox is ongoing, as is the application of deep learning techniques to potentially improve prediction performance by allowing task-specific features to be learned.