An Auto-Encoder Strategy for Adaptive Image Segmentation Evan M. - - PowerPoint PPT Presentation

▶
an auto encoder strategy for adaptive image segmentation
SMART_READER_LITE
LIVE PREVIEW

An Auto-Encoder Strategy for Adaptive Image Segmentation Evan M. - - PowerPoint PPT Presentation

Jan 24, 2023 451 likes •609 views

An Auto-Encoder Strategy for Adaptive Image Segmentation Evan M. Yu, Juan Eugenio Iglesias, Adrian V. Dalca, Mert R. Sabuncu Challenge Annotations costs time, money and requires expertise Weeks to manually label a dataset Growing

slide-1
SLIDE 1

An Auto-Encoder Strategy for Adaptive Image Segmentation

Evan M. Yu, Juan Eugenio Iglesias, Adrian V. Dalca, Mert R. Sabuncu

slide-2
SLIDE 2

Challenge

  • Annotations costs time, money and requires expertise
  • Weeks to manually label a dataset
  • Growing segmentation protocol or imaging technology
  • Objective: Segmentation framework with one manual

segmentations or labels

Figure 1: Structural brain MRI and its delineation

1

slide-3
SLIDE 3

Setup

  • Consider a dataset of N MRI scans {x(i)}N

i=1

  • Let s be latent segmentation
  • By Bayes’ rule:

log p(x(i)) = log

  • s

p(x(i)|s)p(s), (1)

  • Evidence Lower Bound (ELBO):

log p(x(i)) ≥ − KL(q(s|x(i))||p(s)) + E

s∼q(s|x(i))

  • log p(x(i)|s)
  • .

(2)

2

slide-4
SLIDE 4

Segmentation Autoencoder (SAE)

  • Variational Autoencoder (VAE)

L = KL(qφ(s|x(i))||p(s)) − E

s∼qφ(s|x(i))

  • log pθ(x(i)|s)
  • .

(3)

  • Typical VAE uses representation s that is typically continuous
  • Our model maps s to a semantic meaningful representation:

qφ(s|x(i)) =

V

  • j=1

Cat(sj|x(i), φ). (4)

  • Likelihood:

pθ(x|s) =

V

  • j=1

N(x; ˆ xj(s; θ), σ2). (5)

  • Spatial Prior

pspatial(s) =

V

  • j=1

pj(sj). (6)

3

slide-5
SLIDE 5

Architecture

4

slide-6
SLIDE 6

Architecture

1 4 8 16 32 32 64 32 16 14 14

5

slide-7
SLIDE 7

Architecture

1 4 8 16 32 32 64 32 16 14 14

6

slide-8
SLIDE 8

Architecture

1 4 8 16 32 32 64 32 16 14 14 14 1

7

slide-9
SLIDE 9

Architecture

1 4 8 16 32 32 64 32 16 14 14 14 1

8

slide-10
SLIDE 10

Architecture

1 4 8 16 32 32 64 32 16 14 14 14 1

9

slide-11
SLIDE 11

Architecture

1 4 8 16 32 32 64 32 16 14 14 14 1

10

slide-12
SLIDE 12

Evaluation

  • Buckner dataset
  • T1 MRI scans and 12 manual labels
  • 1 probabilistic label atlas
  • 30 training subjects and 8 testing subjects
  • Repeated the experiment 5 times with different random subject

assignments to the train/test partitioning.

11

slide-13
SLIDE 13

Qualitative Results

Figure 2: Representative segmentation results obtained with SAE (w/ MRF)

  • n two subjects.

12

slide-14
SLIDE 14

Quantitative Results

Performance Measure Model Haussdorff (mm) Dice Overlap (%) Baseline 3.50±0.06 71.45±0.65 EM Baseline 2.65±0.05 79.70±0.54 SAE (w/o MRF) 2.73±0.04 79.94±0.34 SAE (w MRF) 2.68±0.05 80.54±0.36 Supervised 2.23±0.07 84.60±0.26

Table 1: Mean performance of all methods with their standard errors.

13

slide-15
SLIDE 15

Thank You

More experiments + Implementation: https://github.com/evanmy/sae

14