← Lectures·L6

Cerebrovascular Imaging & Research

Bernal · 1 concepts · 7 questions

Concept 1 / 1

Key Concepts to Memorize

The brain's resource demands:


  • Only 2% of body weight but consumes a disproportionately high share of O2 and nutrients
  • Requires a continuous supply of blood (via arteries); waste removed via veins
  • Blood circulation serves two roles: (A) Delivery of O2/nutrients, (B) Clearance of waste (amyloid, tau, CO2)

Cerebral Small Vessel Disease (cSVD):


  • = damage to small blood vessels in the brain
  • 5 MRI markers of cSVD:

1. White matter hyperintensities

2. Lacunes

3. Microbleeds

4. Perivascular spaces

5. Atrophy

  • Prevalence: affects 90% of subjects in late life, 20–50% in midlife
  • Clinical impact:

- Present in 50% of dementia cases (incl. Alzheimer's)

- Primary contributor to vascular dementia (2nd most common dementia)

- Behind 80% of intracerebral haemorrhages in late life

- Cause of 25% of ischaemic strokes worldwide


Key concept: asymptomatic → tipping point → symptomatic


  • cSVD damage accumulates silently (tolerable) until a tipping point → stroke or dementia (intolerable)
  • Goal: detect and intervene _before_ the tipping point

CIR Lab research focus (2 pillars):


1. Quantification — AI-powered tools to describe cSVD quantitatively

2. Modelling — disentangle cSVD complexity; phenotyping, staging, clinical implications


Supervised vs. Unsupervised Learning (Bernal's AI intro):


  • Supervised: find a function that maps input features → labels (e.g., classify animal species)
  • Unsupervised: find groups based on similarities in input features, with no labels

Domain Randomisation (key specific project):


  • Problem: not enough real patient data + not enough variability → model doesn't generalize
  • Solution: simulate your own training data with random variations
  • Particularly useful when shapes are somewhat consistent across subjects (e.g., brain lesions)
  • Advantages: no need for real patient data or manual annotations; train with many varied images
  • Disadvantage: the data generation model must be realistic enough — if it's too far from real-world data, the trained model won't transfer