1 of 24

Uncertainty

1

2 of 24

Objective

2

Understand the importance of uncertainty modelling

Be aware of possible sources of uncertainty

Know of applications in which cases modelling of uncertainty can be useful

Know of classical ways of assessing uncertainty

Understand the needs for sampling methods and approximations

3 of 24

Why need for uncertainty assessment

3

Knowing limits of prediction

Likelihood to be wrong

Course of action

When to abstain from using result

Applications:

- Diagnostic -> impact on treatment

- Segmentation -> Surgical planning

4 of 24

Example diagnosis course

4

https://pubmed.ncbi.nlm.nih.gov/37821686/

5 of 24

Labelling uncertainty

5

6 of 24

Artefacts and uncertainty

6

https://arxiv.org/pdf/2109.02413.pdf

7 of 24

Different types of uncertainty

7

Aleatoric

    • From randomness
    • Cannot be tackled by more data
    • Example sources: measurement noise, annotation uncertainty

Epistemic

    • From lack of knowledge on data generation
    • More data helps
    • Uncertainty over model and parameters

8 of 24

Subtypes of aleatoric uncertainty

8

Homoscedastic

    • Same for all input data

Heteroscedastic

    • Varies according to data

9 of 24

Ensembling

9

Bootstrap / Bootstrap aggregation

Take N samples with replacement

Evaluate statistics of interest for these N samples / train model for these N samples

Do this k times

Average across the k trials / Calculate variance / confidence interval of estimate across k trials

Typical bagging algorithm:

Random Forests!!

10 of 24

Modeling / sensing of uncertainty

10

Considering model parameters as samples of a distribution and evaluating output according to these samples

Assessing model variability – different outputs based on different inputs

Assessing risk associated with predicted probability

Considering different models and sampling outputs from these models

11 of 24

Bayesian modelling

11

Model parameters are considered as random variables

Data helps refining the parameter distribution

New data distribution given training data

Analytically difficult

12 of 24

Bayesian linear regression

12

https://www.inf.ed.ac.uk/teaching/courses/mlpr/2016/notes/w7b_gaussian_processes.html

13 of 24

Bayesian regression and uncertainty in diffusion MRI

13

https://www.sciencedirect.com/science/article/pii/S1053811918302696?casa_token=dQVirkrBE8wAAAAA:AOdtRHpC0Ed8XaNtEohLYgjDmTDH7jdhYcFcVw07kdt6EnrpP-zKOpChWq6Jw9HZGAxtfXY1No0

Residual variance to assess uncertainty of prediction

Used for weighting samples (subject and voxel)

14 of 24

From Bayesian regression to Gaussian processes

14

15 of 24

Gaussian processes

15

Distribution over functions

Assumption of continuity

If x close to x’, f(x) close to f(x’)

Infinite multivariate Gaussian distribution over function values

Entirely defined by mean and covariance

Covariance symmetric definite positive

16 of 24

Gaussian processes

16

f(x)

x

x1 x2

x3

f1

f

2

f3

x

x1 x2

x3

f(x)

f3

f2 f1

17 of 24

New sample prediction / multiple samples

17

18 of 24

Crucial choice of covariance matrix

18

19 of 24

Possibility to optimize using training data

19

20 of 24

Gaussian processes and associated uncertainty in medical imaging

20

Age prediction and identification of outliers

Uncertainty as out-of-distribution identifier

https://www.sciencedirect.com/science/article/pii/S1053811918302854?casa_token=7Tpl3XhCVakAAAAA:8hTlEetvojqdfEZfPCwmwpzTft6BsqGDeld99hA98N9CYSlPvUOtdaTM6xSfh-nkkihmDHvBXHY

21 of 24

Gaussian process – Prediction of future

21

https://link.springer.com/chapter/10.1007/978-3-319-19992-4_49

22 of 24

Variational inference

22

Evidence lower bound - ELBO

Use of factorization for simplification

23 of 24

Sampling

23

Used for large multi-dimensional, complex to model distribution.

Create approximation of distribution to be used for generation of integral statistics (mean, variance)

No need for knowledge of normalization factor

Ex: Metropolitan Hastings

  1. Choose random sample
  2. Find next sample based on Gaussian distribution centered in
  3. Calculate ratio between target density in and

  • Generate random number
  • If  accept  as
  • Else

24 of 24

Summary

24