Texture Discretization Sensitivity
Sweep GLCM bin width on a synthetic textured lesion and watch contrast, energy and entropy shift by large factors even though the tissue never changes.
- GLCM
- discretization
- IBSI
- texture
- gray-level binning
Gray levels used
18.000
Contrast
4.701
Energy
0.016
Entropy
4.420
Discretized phantom
Continuous phantom (reference)
Feature value vs. bin width
ContrastEntropyEnergy ×20
Sensitivity across the bin-width sweep
| Bin width | Gray levels | Contrast | Energy | Entropy | Homogeneity |
|---|---|---|---|---|---|
| 1.000 | 64.000 | 48.862 | 0.002 | 6.783 | 0.252 |
| 2.000 | 44.000 | 23.398 | 0.003 | 6.140 | 0.315 |
| 3.000 | 30.000 | 11.165 | 0.006 | 5.391 | 0.389 |
| 5.000 | 18.000 | 4.701 | 0.016 | 4.420 | 0.493 |
| 8.000 | 12.000 | 2.287 | 0.039 | 3.519 | 0.598 |
| 12.000 | 9.000 | 1.254 | 0.080 | 2.812 | 0.692 |
| 16.000 | 7.000 | 0.937 | 0.130 | 2.312 | 0.750 |
| 20.000 | 6.000 | 0.964 | 0.193 | 1.965 | 0.790 |
| 25.000 | 5.000 | 0.696 | 0.271 | 1.666 | 0.829 |
| 32.000 | 4.000 | 0.879 | 0.303 | 1.445 | 0.837 |
Texture features are not properties of the tumor alone
The lesion never changes across this sweep — only the discretization does. Contrast and entropy shift by large factors as bin width changes because they are computed on the quantized image, not the physical intensity field. IBSI-style reporting requires fixing the bin width (or the number of gray levels) and stating it explicitly; comparing "contrast" values computed at different bin widths compares different quantities that happen to share a name.