Practical AI Podcast — Episode Summary

Episode: "Seeing Beyond the Scan in Neuroimaging"

Date: April 30, 2025
Hosts: Daniel Whitenack (PredictionGuard), Chris Benson (Lockheed Martin)
Guest: Prof. Gavin Winston (Queen’s University, Neuroimaging/ML Expert)

1. Overview

This episode delves into how artificial intelligence (AI) and machine learning (ML) intersect with neuroimaging—the science of scanning and understanding the human brain’s structure and function. Prof. Gavin Winston discusses the evolution of neuroimaging technologies, the types of data these modalities produce, how AI/ML is transforming diagnostics and workflow, and the challenges unique to applying AI in the clinical context, including data scarcity, explainability, and the human-AI collaboration.

2. Key Discussion Points & Insights

A. What is Neuroimaging? (02:17–03:23)

• Broad field covering various techniques (MRI, CT, PET) to examine brain structure and function.
• Main clinical use: MRI scans to detect abnormalities like epilepsy, tumors, or stroke.
• “Essentially they're just techniques to look into the brain from the outside and try and give us an idea of the structure and function and how these might be altered.”
  — Gavin Winston (02:17)

B. Historical Context: AI & Neuroimaging (03:23–06:02)

• The 1970s advent of advanced imaging (CT, MRI) generated massive data volumes.
• Human ability limited by scale—AI/ML employed to automate analysis, detect subtle patterns, reduce workload.

C. Real-World Medical Practice (06:02–07:26)

• AI/ML’s uptake in clinical practice is slow due to:
  • Ethical concerns
  • Data quality/access
  • Gap between technical potential and real-world application
• "There's a fairly big chasm between what's possible technically versus what is actually used in practice."
  — Gavin Winston (06:24)

D. How Humans Analyze Scans (07:26–10:30)

• Physicians (radiologists/neurologists) combine clinical info with scan images.
• Diagnosis involves pattern recognition and experience—often difficult to articulate an explicit visual “rule.”
• Structure vs. function: Most clinical imaging is structural, mapping function (e.g., language, visual pathways) reserved for pre-surgical planning.

E. Brain Complexity vs. Neural Networks (12:05–14:35)

• Biological brains are orders of magnitude more complex than current artificial neural networks.
• "There's a big gap between the level of complexity of what we simulate and what we actually are doing in our own brains."
  — Gavin Winston (13:21)

F. Applying ML to Neuroimaging Tasks (15:04–19:32)

• Categorization of ML tasks:
  • Classification (disease presence/absence)
  • Localization (which brain area is affected)
  • Prognosis (predict outcomes, tailor treatments)
• Frequently used approaches:
  • Support Vector Machines (SVMs) for simpler features
  • Convolutional Neural Networks (CNNs) for 3D image data
• Key insight: Most current ML papers tackle binary diseases (“disease X or not?”), but real-life diagnostics often need more nuanced, multi-class decisions.

G. Data Challenges in Medical AI (19:32–22:42)

• Scarcity of high-quality, labeled data: Studies might have <100 cases, tiny compared to LLM training corpora.
• Labeling (e.g., drawing affected areas) is labor-intensive and subject to inter-rater variability.
• Data augmentation (modifying existing samples) is possible, but fully synthetic neuroimaging data isn’t robustly used.

H. Explainability and Clinical Trust (23:27–26:45)

• Physicians require transparency to trust AI outputs, especially for high-stakes, life-impacting decisions.
• “A lot of the AI algorithms could appear to be like a black box … they want to know what's happening, so they can trust the algorithm.”
  — Gavin Winston (24:33)
• Mention of the MELD Study: Outputs not only suggest likely abnormal brain regions but also the features influencing that prediction, facilitating clinician acceptance.

I. Barriers to Adoption — Cultural, Ethical, Legal (26:45–37:29)

• Ethical/data privacy and legal frameworks lag behind research advances.
• Patient data security, jurisdictional laws, and professional culture all influence deployment rate.
• “There's definitely a lot of scope to working out all of these legal issues and how they best [are] dealt with.” — Gavin Winston (37:06)

J. AI in the Diagnostic Workflow (30:08–34:41)

• Human expertise remains critical, especially for subtle or ambiguous abnormalities.
• ML excels at pattern recognition and can pre-screen or prioritize scans, bringing potential abnormalities to the forefront.
• “It's going to augment the abilities … more efficient, more smooth … but it's never going to completely replace the human aspect.”
  — Gavin Winston (33:31–34:41)
• Major limitations today: false positives, imperfect accuracy—human judgment essential.

K. The Future & Aspirational Use Cases (39:07–41:36)

• Vision: Automated triage—scans prioritized for review by degree of abnormality likelihood.
• ML-generated reports and visual annotations augment radiologist reading.
• AI can alert treating physicians to critical findings faster, improving patient outcomes.

3. Notable Quotes & Memorable Moments

• “When we have neural networks … you don't realize just how complicated the brain is, just how many billions of neurons it has, and how they're all vastly interconnected.”
  — Gavin Winston (13:21)
• “One of the biggest challenges is having access to the data in the first place … 100 subjects to someone in machine learning, that is a minuscule number.”
  — Gavin Winston (20:23)
• “What I would like to see is as a radiologist sits down to report the 20 scans they've got, instead of scan 1 to 20, it's scan most likely to be abnormal to least likely to be abnormal.”
  — Gavin Winston (39:59)

4. Timestamps for Important Segments

• 02:17 — What is neuroimaging and why it matters
• 03:23 — Evolution of neuroimaging & AI’s arrival
• 06:24 — Real-world barriers to AI in clinical settings
• 10:30 — How structure and function relate in diagnosis
• 13:21 — Neural networks vs. actual brain complexity
• 15:04 — Key ML applications in neuroimaging
• 19:32 — Unique technical/data challenges in the field
• 23:27 — Explainability and the MELD study
• 30:08 — Performance comparisons: Human vs. AI
• 33:31 — The role of AI: Augmentation, not replacement
• 39:59 — Vision for AI-powered diagnostic workflow

5. Tone, Language, and Takeaways

• Conversational yet deeply insightful; practical examples and open about current challenges
• Emphasizes careful adoption, clinical context, and partnership between technology and health professionals
• Recurring theme: AI as augmentation rather than replacement; crucial for improving patient care, not eliminating the clinician’s role

This episode is a must-listen for anyone interested in the intersection of medical imaging, AI, and real-world clinical practice—providing both technical insights and a grounded view on translation from lab to hospital.