Innovate UK Call AI in Digital Pathology · Data Controller Algorithm Caldicott Anonymizing Data...
Transcript of Innovate UK Call AI in Digital Pathology · Data Controller Algorithm Caldicott Anonymizing Data...
iCAIRD
iCAIRD
Industrial Centre for Artificial Intelligence Research in Digital Diagnostics
June, 2019
Cambridge, MA
Santa Clara, CA
Peter Hamilton Leader Image Analytics, Philips Digital Pathology Solutions, Hon Professor of Tissue Imaging, QUB, Belfast
Andy Smout Vice President Research, Canon Medical, Edinburgh
Colin McCowan Professor of Health Informatics, University of Glasgow and Glasgow Safe Haven
Alison Murray Professor of Radiology, Director of SINAPSE, University of Aberdeen, NHS Grampian
David Harrison NHS Lothian & Universities of St Andrews, Edinburgh & Glasgow
Key features • Tried & tested partnerships •NHS NSS Board • Safe havens & HDRUK • SINAPSE •National PACS •Direct link to clinicians
Imaging Centre of Excellence
2
Democratising AI: Reducing Barriers to Entry
1. The Domain Barrier
2. The Data and Annotation Barrier
3. The Clinical Validation Barrier
4. The Regulatory Barrier
5. The Channel to Market Barrier
Without an existing product line and an established clinical collaborator network it is hard for SMEs to know where to focus
Machine learning solutions require huge amounts of data to generalise well. It is hard for SMEs to get access to that scale of data and harder still to annotate it accurately
Without a product already integrated into the clinical workflow it is difficult for SMEs to validate algorithms in a real-world multi-centre setting and generate the evidence needed to demonstrate their clinical effectiveness
Healthcare AI has stringent requirements on safety and effectiveness. These can daunt SMEs wanting to enter the market
Without an established global sales and marketing organisation it is difficult for SMEs to access a large enough customer base, and without an established reputation it is equally hard to form commercial partnerships with established vendors
3
The Safe Haven AI Platform (SHAIP)
Healthcare Enterprise Technology Company
WORKSPACE
Data Controller
Algorithm
Caldicott
Anonymizing
Data Portal
Machine
Learning Portal Safe Haven
Researcher
Data Scientist
Clinician 1: Researcher works with Clinician to
identify a potential new AI algorithm 2: Clinician identifies a suitable
cohort of patients for research
6: Researcher uses anonymizing data
portal to explore data and generate
ground truth without encountering PHI
7: Data scientist uses machine learning
portal to train new algorithm
3: Caldicott guardians approve use of
data from cohort for specified research
5: Data approved for research is pulled from
clinical systems and cached in the workspace
4: Data controller
allocates cohort to
company workspace
4
Philips-centric pathology AI Exemplars: transforming pathology, enabling pathologists
5
Endometrial AI Pathology App
Why? • 42% of gynaecological specimens are endometrial • uniform with >95% comprising single slide • Exclusion of neoplasia is key pivot • Only 3% of endometrial biopsies show adenocarcinoma • Only 1.5% are atypical • >95% of biopsies are benign Perfect setting to develop AI to screen out non-malignant/ atypical cases and reduce NHS workload Technically challenging Benign patterns show considerable heterogeneity in pattern due to endogenous and exogenous hormonal influence.
Cervical AI Pathology App
Why? • 26% of gynaecological specimens are cervical biopsies
(including punch biopsies, polyps and LLETZ/LOOP excisions)
• The primary reason for a cervical biopsy is for the assessment of cervical intra-epithelial neoplasia (CIN) and exclusion of invasive squamous or adenocarcinoma.
Perfect setting to develop AI to identify invasive cancer, generate automated reports and reduce NHS workload Technically challenging Requires contextual image mapping at multiple resolutions to distinguish CIN from Invasive cancer particularly the identification of microinvasion
If these targets are achieved there would be an 85% time saving in consultant time across these specimen types. This would result in a saving of £185,650 per annum for NHS GGC which is 54% of reporting time in gynaecological pathology. Extrapolated across the UK, this would equate to a saving of £9.3 M per annum
6
Sustainability & SME growth
7
Market reach
Public confidence
Sales
Regulatory
Accelerator
NSS NHS procurement
Health economics
UK SME ecosystem
SME Application SME
Engagement Team
Searchable data lake access for R&D and product development
Conduit to pathologists for application development and annotation
Access to data scientists for deep learning expertise in pathology
Use of validated tools for fast track deep learning development
Conduit to established industry platform as an option to accelerate pace to market
Entry to Accelerator Programme for training, mentorship and leadership in health-tech
Access to interdisciplinary team of health, technology and industry experts
On-line educational programme for pathologists and data scientists, business and innovators
SME Engagement
Digital pathology
8
Implementation Economic case unproven Early adopter risk & competition Interoperability Artificial intelligence Platform and apps or tied to hardware provider Future of pathology High volume High complexity
Window of opportunity
• Clinical implementation
• Operating system for artificial intelligence apps & interoperability
• High quality artificial intelligence
• Clinical trials exemplars
• Open source datalake 9
• Clinical implementation – Visiopharm, Definiens, Indica
• iCAIRD funding – Blackford Analytics, Glencoe, EPCC
• Clinical trials exemplars – kidney cancer, links to industry, cancer centres & tissue, attracting in clinical trials, CSO Innovation Fellows
• Open source datalake & interoperability
10