Table 1 Quantitative evaluation of the image fidelity and diversity of synthetic samples

From: Realistic morphology-preserving generative modelling of the brain

Dataset

Model

FIDa,b

MMD

MS-SSIMc

4-G-SSIM

UKB

Ours

0.0026

9.53 × 10−7

0.67 ± 0.05

0.40 ± 0.02

UKB

HA-GAN12

0.0047

1.4 × 10−6

0.66 ± 0.08

0.42 ± 0.06

UKB

CCE-GAN10

0.0888

4.08 × 10−5

0.66 ± 0.05

0.38 ± 0.01

UKB

LS-GAN34

0.0171

1.26 × 10−6

0.65 ± 0.04

0.38 ± 0.01

ADNI

Ours

0.0075

5.21 × 10−7

0.69 ± 0.07

0.40 ± 0.06

ADNI

HA-GAN12

0.0219

1.99 × 10−6

0.59 ± 0.09

0.38 ± 0.06

ADNI

CCE-GAN10

0.1680

5.96 × 10−5

0.59 ± 0.06

0.37 ± 0.03

ADNI

LS-GAN34

0.0763

8.86 × 10−6

0.59 ± 0.05

0.37 ± 0.2

  1. aA lower FID and MMD indicate better distribution alignment between the real and synthetic samples, while a lower MS-SSIM and 4-G-SSIM indicate higher diversity of synthetic samples.
  2. bFID and MMD are lower bounded by 0 with no upper bound while MS-SSIM and 4-G-SSIM are lower bounded by 0 with an upper bound of 1.
  3. cFor MS-SSIM and 4-G-SSIM the mean and standard deviations are presented.
Back to article page