Skip to main content

Table 1 Overview of perturbations applied to simulate multi-center data. Perturbation types and ranges vary with labels and MRI sequences according to the recommendation in [8]

From: Radiomics for glioblastoma survival analysis in pre-operative MRI: exploring feature robustness, class boundaries, and machine learning techniques

Perturbation

Sequences

Segmentaion Labels

Perturbation assessment or range

Number of perturbations per sequence/label

Additivie Gaussian noise

All

Match SNR range

10 per sequence

Histogram binning

All

PyRadiomics recommended range (30–130 bins)

54 per sequence

Inter-rater simulation

All

Match Inter-rater DICE in [18]

10 per label

Voxel size

All

1–1.5 mm, isotropically

10 per sequence

Slice spacing

FLAIR, T2

1-4 mm, imaging recommendations in [8]

10 per sequence

 

T1, T1c

1–1.5 mm, imaging recommendations in [8]

10 per sequence

K-space subsampling

All

Random subsampling, factor 0.8–1.0

21 per sequence