Generative AI in MS Neuroimaging

Delivered: 15 May 2025
Event: NMSS
Format: Invited lecture
Source deck: 20 slides

From the slides

Deck notes

This lecture surveyed how generative models may help MS neuroimaging, from synthetic contrasts and privacy-preserving data to radiology reports. The same deck was used for an IPMSA session later in 2025.

The question

Can a model that learns the distribution of MRI data extend what researchers can study without introducing new error or bias?

Deck outline

The distinction between discriminative and generative models
Synthetic MRI contrasts and enrichment of clinical scans
Synthetic data as a route to greater privacy
Generative models for treatment-effect research
Bias, fairness and equitable model performance
Small samples, limited validation and the gap to clinical impact

Take-away

Generative AI is versatile, but clinical value still depends on external validation, meaningful outcomes and careful assessment of fairness.

Original deck

Slide by slide

Read the 20-slide presentation in order, or use the index to jump.

Slide 01
Generative AI in MS Neuroimaging
Open full-size slide
Arman Eshaghi, MD, PhD
NIHR Advanced Fellow
Queen Square Institute of Neurology
University College London
United Kingdom
Slide 02
Disclosures
Open full-size slide
Received research grants from the Medical Research Council (MRC), National Institute for Health and Social Care Research (NIHR), Innovate UK, Biogen, Merck, and Roche.
Received Speaker’s honoraria from Roche.
He served in advisory board meetings of Merck Serono and Bristol Myers Squibb.
He is the founder and equity stakeholder in Queen Square Analytics Limited.
He serves on the editorial board of Neurology (American Academy of Neurology).
Slide 03
Outline
Open full-size slide
Definition
Applications
Limitations and future
Speaker notes
Routine-care MRI and advanced MRI
Slide 04
Definition
Open full-size slide
AI
Machine learning
Deep learning
Slide 05
Definition
Open full-size slide
Discriminative
Generative
Speaker notes
Discriminative: “We learn a boundary.” Generative: “We learn the world of the data:”
Slide 06
History
Open full-size slide
Slide 07
Applications: generative machine learning
Open full-size slide
Understanding
Enriching clinical data
Prescription
Equitability
Automating tasks
Disease course
Synthetic data
Treatment effect
Fairness
Radiology report generation
Privacy preserving
Bias
Q&A with persons with MS
Slide 08
Understanding disease course
Open full-size slide
Eshaghi et al, 2021, Nature Communications
Ganjgahi et al, 2025, Nature Medicine
References
- Eshaghi et al, 2021, Nature Communications — PubMed search
- Ganjgahi et al, 2025, Nature Medicine — PubMed
Slide 09
Enriching clinical MRIs
Open full-size slide
Generative model
T2
FLAIR
T1
Synthetic contrast
T1-weighted MRI
Narayana et al, Radiology, 2019
Speaker notes
Sensitivity of 0.72 and specificity of 0.75
References
- Narayana et al, Radiology, 2019 — PubMed
Slide 10
Synthetic data to enrich clinical MRIs
Open full-size slide
Goebl et al,
Nature Communications, 2025
Slide 11
Generative models: Synthetic data for privacy
Open full-size slide
Bad for privacy
Good for privacy
Ziller et al, Nature Machine Intelligence, 2024
Speaker notes
smaller → stronger privacy smaller 𝜀→ stronger privacy
References
- Ziller et al, Nature Machine Intelligence, 2024 — PubMed search
Slide 12
Prescription: treatment effects
Open full-size slide
Fallet et al, 2022, Nature Communications
References
- Falet et al, Nature Communications, 2022 — PubMed
Slide 13
Bias and Fairness
Open full-size slide
Seyyed-Kalantari et al, 2021, Nature Medicine
Speaker notes
a, The underdiagnosis rate, as measured by the no finding FPR, in the indicated patient subpopulations. b, Intersectional underdiagnosis rates for female patients (b(i)), patients aged 0–20 years (b(ii)), Black patients (b(iii)), and patients with Medicaid (b(iv)). c,d, The overdiagnosis rate, as measured by the no finding FNR in the same patient subpopulations as in a and b. The results are averaged over five trained models with different random seeds on the same train–validation–test splits. 95% confidence intervals are shown. Subgroups with too few members to be studied reliably (≤15) are labeled in gray text and the results for these subgroups are omitted. Data for the Medicare subgroup are also omitted, given that data for this subgroup are highly confounded by patient age. The “no finding” FPR (underdiagnosis rate) is highest for female, younger (0–20), Black, Hispanic, and Medicaid-insured patients—and it worsens at intersections (e.g., Black female, female + 0–20, 0–20 + Medicaid). In short: the groups already at risk in care are the most likely to be told “nothing is wrong.”
References
- Seyyed-Kalantari et al, 2021, Nature Medicine — PubMed
Slide 14
Bias and Fairness
Open full-size slide
Langer-Gould et al, Neurology, 2022
References
- Langer-Gould et al, Neurology, 2022 — PubMed search
Slide 15
Equitability: Fairness
Open full-size slide
IS = Inception Score
(Quality + Diversity)
Luo et al, 2025, Science Advances
References
- Luo et al, 2025, Science Advances — PubMed search
Slide 16
Automating tasks: radiology report generation
Open full-size slide
Tanno et al, 2025, Nature Medicine
References
- Tanno et al, 2025, Nature Medicine — PubMed
Slide 17
Slide 17
Open full-size slide
Slide 18
Limitations: generative models in MRI
Open full-size slide
Minimal clinical translation
- Small samples
- Lack of prospective and external validation
Lack of clinical meaningfulness and cost-effectiveness
- Clinical outcome
Slide 19
Summary
Open full-size slide
Generative models and their history
Wide-ranging research applications
Limited clinical impact (so far) but promising
Fundamentals to improve research capabilities
- Outcomes
- High-quality study design
Slide 20
Slide 20
Open full-size slide
Email: a.eshaghi@ucl.ac.uk