A learning strategy for contrast-agnostic segmentation of brain MRI - - PowerPoint PPT Presentation

▶
a learning strategy for contrast agnostic segmentation of
SMART_READER_LITE
LIVE PREVIEW

A learning strategy for contrast-agnostic segmentation of brain MRI - - PowerPoint PPT Presentation

Aug 21, 2023 267 likes •531 views

A learning strategy for contrast-agnostic segmentation of brain MRI scans Benjamin Billot Billot 1 , Greve 2 , Van Leemput 2 , Fischl 2 , Iglesias* 1,2,3 , Dalca* 2,3 1 Centre for Medical Image Computing, UCL 2 Martinos Center for Biomedical

slide-1
SLIDE 1

A learning strategy for contrast-agnostic segmentation of brain MRI scans

Benjamin Billot

Billot1, Greve2, Van Leemput2, Fischl2, Iglesias*1,2,3, Dalca*2,3

1Centre for Medical Image Computing, UCL 2Martinos Center for Biomedical Imaging, Massachusetts General Hospital 3Computer Science and Artificial Intelligence Laboratory, MIT

*contributed equally

benjamin.billot.18@ucl.ac.uk

slide-2
SLIDE 2

2

Segmentation

Segmentation method

slide-3
SLIDE 3

3

Types of methods

Methods Speed Manual

  • +++

Multi-atlas segmentation

  • +

Bayesian segmentation + ++ Supervised CNN +++

  • Modality-agnostic
slide-4
SLIDE 4

4

Modality-specific CNN

T1 Supervised CNN

slide-5
SLIDE 5

5

Supervised segmentation

T1 Supervised CNN T2 Supervised CNN T1 + T2 Supervised CNN

slide-6
SLIDE 6

6

Problems with supervised CNNs

  • only work on modalities they were trained with
  • require supervised data
  • sensitive to pre-processing
slide-7
SLIDE 7

7

Solution: Synthesise data…

CNN trained with synthetic data

slide-8
SLIDE 8

8

…of random contrast !

Supervised CNN Set of anatomical segmentations

SynthSeg

slide-9
SLIDE 9

9

Outline

Introduction Methods

  • Generative model
  • Training

Experiments and results

  • Experimental set-up
  • Results

Conclusion

slide-10
SLIDE 10

10

Generation of T1 contrast

Label map Spatial deformation GMM sampling Blurring Bias field

slide-11
SLIDE 11

11

Generation of random contrast

Label map Spatial deformation GMM sampling Blurring Bias field

slide-12
SLIDE 12

12

Predicted label map Image generative model Modality agnostic CNN Average Dice loss Data generation

SynthSeg training overview

Generated image Deformed label map Label map Backpropagation

slide-13
SLIDE 13

13

Outline

Introduction Methods

  • Generative model
  • Training

Experiments and results

  • Experimental set-up
  • Results

Conclusion

slide-14
SLIDE 14

14

ABIDE ADHD HABRE GSP MCIC OASIS PPMI DBS T2 T1 PD T1 Testing

T1-39:

39 subjects

T1mix:

1,000 subjects

FSM:

18 subjects

T1-PD-8:

8 subjects

Training

Datasets

slide-15
SLIDE 15

15

Competing methods

  • T1-baseline: T1 supervised CNN
  • SAMSEG [1]: modality-agnostic Bayesian segmentation
  • SynthSeg
  • SynthSeg-rule: trained with realistic contrasts

[1] Puonti et al., Neuroimage, 2016.

slide-16
SLIDE 16

16

Dice scores

T1mix

slide-17
SLIDE 17

17

Dice scores

T1mix

slide-18
SLIDE 18

18

Dice scores

T1mix

slide-19
SLIDE 19

19

Dice scores

T1mix

slide-20
SLIDE 20

20

T1-39 T1-FSM Ground Truth T1 baseline SAMSEG SynthSeg-rule SynthSeg

T1 segmentation examples

slide-21
SLIDE 21

21

T2-FSM PD-PD8 Ground Truth T1 baseline SAMSEG SynthSeg-rule SynthSeg

T2-PD segmentation examples

N/A N/A

slide-22
SLIDE 22

22

Key points

  • SynthSeg enables fast contrast-agnostic segmentation of

brain MRI scans, without retraining.

  • SynthSeg does not require any preprocessing.
  • SynthSeg only requires a set of segmentations as training data.
  • Augmentation beyond realistic measures enables better

generalisation.

slide-23
SLIDE 23

23

Future directions

SynthSeg PV-SynthSeg

slide-24
SLIDE 24

24

Acknowledgments

Funding: Collaborators:

slide-25
SLIDE 25

25

Useful links

  • A Learning Strategy for Contrast-agnostic MRI Segmentation

MIDL 2020 https://arxiv.org/abs/2003.01995

  • Generative model:

https://github.com/BBillot/lab2im

  • SynthSeg:

https://github.com/BBillot/SynthSeg

  • Partial Volume Segmentation of Brain MRI Scans of any Resolution

and Contrast MICCAI 2020 https://arxiv.org/abs/2003.01995