London Medical Imaging & Artificial Intelligence Centre for Value-Based Healthcare

Professor Reza Razavi

Centre Director

Unrestricted © Siemens Healthcare GmbH, 2018

Edit “Author | Department” in the “Header and footer” for all pages.

This presentation uses the preinstalled Office font Calibri. If you send it to someone internally or externally, please convert it into a pdf to ensure that the content stays as intended.

Our template contains layout guidelines. Their purpose is to indicate key margins and mandatory areas on slides.

Copy and paste Copy only full slides when they use the same new design. Otherwise the file size will be very large. Use the function Insert > Insert Content > Unformatted Text to paste a copied text. The text will now use the same existing format.

Each presentation needs to be categorized, with the author being responsible for choosing the right category: – Unrestricted – Restricted – Confidential – Strictly confidential The preset default category for presentations is “restricted”. Start > Replace > insert “Restricted ©” and “the new category ©”

• To ensure a clean and swift workflow with bullet points, please use the PRE-SET PLACEHOLDERS or FORMATTED TEXTBOXES – do not use “normal” textboxes that have been added via the steps add textbox. These textboxes cannot be formatted with the automatic formatting step

• AUTOMATIC INDENTATIONS IN PLACEHOLDERS are only to be done using the tool decrease or increase the list level (or Shift + Alt + / )

• Formatted textboxes/ placeholders are available in the template – just make a copy OR: generate a new placeholder by following these steps Start new slide choose layout "Content Slide"

• Placeholders can be filled and then copied. The copied placeholder will keep its formatting

Add a Footnote/Source/ Disclaimer directly into the footer.

No. of patients

No. of procedures

Reimbursement Fee

The traditional healthcare business model is changing

Patient population 4

Patient population 3

Patient population 2

Patient population 1

∑

Inpatient clinics

Outpatient clinics

Other care professions

Lab and pathology

Anesthesiology

Radiology

Intensive care

Emergency care

Neurology

Surgery

Cardiology

Oncology

Fee-for-service

= Health outcomes and costs per patient

Optimisation

criteria

Fee-for-outcome

Oth

er ho

spitals

Prim

ary care

Oth

er rehab

, elder care etc.

Clinical demand definition of patient population

Patient group experts

Information provider, Decision enabler

Functional experts and teams

Other providers

Healthcare costs

HC DI

Page 3 | Unrestricted © Siemens Healthcare GmbH, 2017

1) Source: www.ai-healthandlifesciences.com

50% potential reduction in medical treatment costs1)

Artificial intelligence (AI) for personalized decisions:

Big data support Personalised diagnosis & therapy Creating Benefits for

• Patients: Cost efficient precision medicine • NHS: Cost savings for the NHS • UK: Transformative economic growth By Enabling • Time efficient automated reporting

• Optimal treatment stratification

• Major redesign of clinical pathways to improve outcomes and efficiency

• New business opportunities for UK companies and job creation

50% potential reduction in medical treatment costs1)

Artificial intelligence (AI) for personalized decisions:

Big data support

Our Centre Vision: Driving the Transformation towards

Value-Based Healthcare

Centre Overview

• Focus on 12 clinical pathways for maximum impact

• 10 SMEs core partners supported on site

• Siemens European Stratified Medicine R&D headquarters

• Co-location of academic, clinical and industry/SME staff: “sandpit” for

rapid prototype iteration

• World-leading expertise in imaging, AI, clinical research

• Substantial computational infrastructure, in partnership with IBM & NVIDIA

• “Bringing the algorithms to the data”

• Engagement with NHS commissioners and health economics

AInostics Thames Mammography

Centre Members

St Thomas’ Hospital Hub

Radiologists/clinicians & health economics Researchers in advanced imaging & AI

Multinational Industry (e.g. Siemens) Small & Medium Enterprises

Radiologists/clinicians & health economics Researchers in advanced imaging & AI

Multinational Industry (e.g. Siemens) Small & Medium Enterprises

Wellcome/EPSRC Centre for Medical

Engineering

Industrial strategy: growing SMEs

SMEs

Clinicians

Health economics

AI/data science experts

Regulatory, QMS & data governance

Computing infrastructure

Siemens, GSK, NVIDIA, IBM

Wellcome/EPSRC Centre

HDR UK LHCRE

BRCs Medical imaging

Business seminars

NHS commissioner engagement

Software development

IP & commercial

strategy

NIHR Med Tech Co-operative

KiTEC The Foundry

Venture Capital

NHS data

Acute Chest Pain

Patient Presents at

A&E with Stable

Chest Pain Clinical Assessment

Non-invasive CT

Coronary

Angiography

Patient Presents at

A&E with Stable

Chest Pain

Admission pending

further investigation

Clinical Assessment

Admission pending

further investigation

Discharge back to

Primary care

Negative

for CAD

CAD

explaining

symptoms

Key Question: Does introducing a CT scan

for low to intermediate risk ACP patients

result in earlier diagnosis and subsequent

appropriate treatment or discharge?

Interesting Facts/Benefits:

• At St Thomas’ 12 patients per day are

admitted for observation with acute chest

pain, at a cost of £1000 per patient.

• Cardiac CT costs £200. 80-90% of

patients could potentially be discharged

earlier with this facility available.

•ACP patients are often discharged to GPs

with no definite diagnosis and are,

subsequently, often referred back to

cardiology.

• NICE acknowledges the use of cardiac CT

• Rapid turnaround of scan and report

within 4 hr wait

Deep learning for automatic reporting of X-rays

Annotation of all images required for training & testing Automated tagging through deep nets for natural language modelling

Initial application: fully-automated reading of chest x-ray images One million exams from Guy & St Thomas’ Archives (2005-15)

Opacity

Pleural Effusion

Enlarged Heart

Enlarged heart

Fully-automated tagging of a raw image done in milliseconds

Semantic segmentation learned from weak labels only

Medical Device

Enlarged heart

Clinical Finding Sensitivity Specificity Accuracy

Medical Device 0.925 0.934 0.929

Opacity 0.883 0.906 0.894

Pleural Effusion 0.951 0.950 0.951

Enlarged Heart 0.917 0.882 0.900

Atelectasis/Collapse 0.852 0.846 0.849

Pneumothorax 0.868 0.909 0.889

Nodule 0.746 0.865 0.806

COPD 0.892 0.865 0.878

Edema 0.912 0.936 0.924

Infection 0.823 0.859 0.841

Abnormal chest radiographs were detected with a precision of 0.94; sensitivity of 0.95 and specificity of 0.71

Pesce E, et al. Radiology 2018 (in print)

Run-time performance: 15 fps

Machine Learning for Organ Recognition in Fetal screening

Baumgartner CF et al IEEE Trans Med Imaging. 2017 Nov;36(11):2204-2215

Deep Learning for Prognostic Modelling in Cancer

For the treatment of oesophageal cancer, neoadjuvant chemotherapy

(NC) is given as a first step before surgery Responders to NC have improved survival after surgery, but prognosis

is worse for non responders

Survival rates in 107 patients

Recent radiomics studies have suggested that texture features extracted from 18F-FDG PET scans can predict NC response

Deep neural networks offer an alternative radiomics strategy: they learn predictive imaging features from raw scans

Responder

Non Responder

Best classifier using >100 texture features achieves 66.8% accuracy A convolutional network fusing adjacent slices reaches 73.4%

Ypsilantis et al. 2015 PLOS One