Skip to main content
. 2025 Jul 16;12:1589587. doi: 10.3389/fmed.2025.1589587

Table 5.

Performance metrics for segmentation of classical texture analysis methods (U-Net, Gray-Level Co-occurrence Matrix, and Local Binary Pattern) evaluated with and without data augmentation on the Polyp dataset.

Model Accuracy (%) Recall (%) Specificity (%) Dice (%) IoU (%)
U-Net (Augmentation) 95.00 99.47 90.00 94.5 ± 0.35 90.2 ± 0.41
U-Net (No Augmentation) 98.00 99.00 98.00 92.3 ± 0.41 87.7 ± 0.46
LBP (Augmentation) 96.50 99.00 89.00 90.1 ± 0.45 84.8 ± 0.51
LBP (No Augmentation) 98.00 99.49 96.00 88.0 ± 0.48 82.3 ± 0.53
GLCM (Augmentation) 94.50 99.47 88.00 86.2 ± 0.50 79.9 ± 0.56

Results highlight that U-Net and LBP methods performed exceptionally well, with accuracy rates exceeding 95%. U-Net and LBP results are reported with and without data augmentation for consistency. Bold values indicate the best results obtained.