Using MRI from Clinical Practice to Inform Prognosis

Artificial Intelligence Applications from Diagnosis to Prognosis

Arman Eshaghi, MD, PhD

NIHR Advanced Fellow

Queen Square Institute of Neurology, UK

Semi-automatic

tools

~2015

Deep learning

Next

Multimodal agentic

reporting

Now: 2026

Early 2000s

MRI in diagnosis

(McDonald criteria)

2010s

Automatic

After 2015

FDA-approved AI

devices for brain MRI

Emerging

Text

generation

with clinician oversight

Single task

c. 2023

Multiple tasks

Capability

Qualitative expert inspection

Quantitative measures

Role of MRI

Injury

×

Reserve

Outcome

time

Clinical examination

Disability (RAW / PIRA)

Age

MRI

OCT

Disease activity

Cognition

CSF

Blood

Quality of life

Tur et al., Brain (2026). doi:10.1093/brain/awag047

Explanatory Power for Physical Ambulatory Disability (EDSS Variance)

Upper Cervical Cord Area (MUCCA / CSA)

[k = 10 studies, N = 1,171]

23.0%

Spinal Cord Lesions (SCLN / SCLV)

[k = 8 studies, N = 948]

14.4%

Normalized Brain Volume (NBV / BPF)

9.0%

Brain T2 Lesions (T2LV / WMLV)

[k = 9 studies, N = 1,085]

6.8%

Meta-analysis (unpublished)

93.3% AI Sensitivity case-level detection

58.3% Routine Reports case-level detection

Baseline FLAIR

Follow-up FLAIR

AI Lesion Overlay

Barnett et al., npj Digital Medicine (2023) 6:196 · doi:10.1038/s41746-023-00940-6

• DIMENSION — RESEARCH — ROUTINE CARE
• Technical — Harmonised — → — Heterogeneous
• Population — Hundreds — → — Millions
• Ethics — Opt-in — → — Opt-out
• Governance — Central — → — Local
• Fairness — Selective — → — Inclusive
• Design — Traditional trials — → — Pragmatic trials

Provides brain regions and lesions from a single T1, T2, FLAIR or PD scan

Goebl et al, Nature Communications, 2025

https://mindglide.ms-pinpoint.com/

Willard et al, 2025, Brain

Preziosa et al, 2026, Brain, doi:10.1093/brain/awag320

Eshaghi et al, 2021, Nature Communications

Ganjgahi et al, 2025, Nature Medicine

Teo JT, Eshaghi A, et al. Nature Reviews Neurology (2026). doi:10.1038/s41582-026-01225-8

Equity

Will patients benefit worldwide?

Bias

Does it work equally well for everyone?

Effectiveness

Does it improve care in practice?

Patient-reported outcomes

Translating AI to practice: empirical evidence across N=220 authorized devices (US FDA · European CE · Asia)

Input modality distribution

Cross-jurisdictional clinical focus

Clinical indications & scope (N=220)

Empirical synthesis from US FDA CDRH databases (510k/De Novo), UK MHRA, EU MDR, and Asian Registries (MFDS, NMPA, PMDA).

• Evidence — Strengths — Weakness — AI clinical maturity
• Cerebral lesions — Proven — Fully automatic — Sensitivity varies across regions — Most mature – regulatory approval
• Brain volumes — Proven — Fully automatic — No clinical consensus — Mature – regulatory approval
• Spinal cord atrophy — Proven — Motor disability — No automatic software — No evidence
• Cord lesions — Proven — Motor disability — No automatic software — No evidence

Learning from routine hospital scans

Newly diagnosed MS: the model highlights the regions behind its ‘multiple sclerosis’ call (illustrative)

5.24 million

MRI and CT volumes, no manual labels

20 years

of routine care, one US health system

156 diagnoses

tested, across scanner makes and field strengths

Kondepudi et al, Nature Medicine, 2026

Start date

Ethics and CAG approval

Sheffield and Coventry started

Closed recruitment

Oct 2022

June 2023

June 2025

Feb 2026

Leeds: 1,100

Pretrain

Learn from routine brain and spinal cord MRI, without manual labels

Sheffield: 3,500

Nottingham: 4,150

Adapt

Fine-tune to brain and spinal cord anatomy and lesions

Cambridge:

1,500

International centres

Coventry:

2,500

Luton &

Dunstable: 1,200

Predict

Disability worsening and MRI-derived subgroups

Swansea: 600

Amsterdam UMC,

the Netherlands

5000

Cardiff: 1,600

UCL: 4,100

Imperial: 2,700

King’s: 3,500

Montreal Neurological

Institute, Canada

1,100

Federated: data stays in each hospital, only model parameters are shared

RESULTS COMING

ECTRIMS Global 2027 · Vienna, Austria

Approximate sample size: 29,000 patients

Blankemeier L, Kumar A, et al. Nature (2026). doi:10.1038/s41586-026-10181-8

MRI: diagnose and monitor.

Adds value for prognostication.

Demonstrate patient benefit.

Decision aids, fairness checks and prospective evidence.

Urgent need for better outcomes.

Potential with technological innovations.

Crossroads of AI and data to revolutionise MS prognostication.

Terry Lock · Trishna Bharadia · Helen King

Thank you

MS-PINPOINT team, hosts and collaborators

MS-PINPOINT TEAM

PPI STEERING COMMITTEE

Patient and public involvement advisers

Jon

Stutters

Philipp

Goebl

Barbara

Brito Vega

Sean

Apap Mangion

Stef

Piatek

HOSTS & COLLABORATORS

Ahmed

Toosy

Daniel

Alexander

Frederik

Barkhof

Wallace

Brownlee

Jeremy

Chataway

Declan

Chard

Olga

Ciccarelli

Floriana

De Angelis

Alan

Thompson

Will

Brown

With Philipp Goebl and Barbara Brito Vega

Alasdair

Coles

Ed

Peake

Gavin

Brittain

David

Paling

Neil

Harrison

Derek

Jones

Emma

Tallantyre

Stuart

Currie

Helen

Ford

Nikos

Evangelou

Radu

Tanasescu

Richard

Nicholas

Dobson

Rhian

Raftopoulos

James

Teo

Owen

Pearson

Rod

Middleton

Tarunya

Arun

Joep

Killestein

Menno

Schoonheim

RA

Eva

Strijbis

Douglas

Arnold

Rozie

Arnaoutelis

Paul

Giacomini

Adil

Harroud

Louis

Collins

Vladimir

Fonov

Sridar

Narayanan

Robin

Wolz

Email: a.eshaghi@ucl.ac.uk