Table 5 The individual contributions of MultiDimensional blocks and encoder-decoder routes are being studied using ablation. Using 5-fold cross-validation, the EndoSLAM findings are obtained
From: Multimodal Biomedical Image Segmentation using Multi-Dimensional U-Convolutional Neural Network
Type | F = 1 | F = 2 | F = 3 | F = 4 | F = 5 | Avg |
|---|---|---|---|---|---|---|
Traditional U-Net | 74.11 | 72.67 | 73.89 | 75.22 | 76.07 | 74.392 |
Only encoder-decoder path | 76.34 | 73.78 | 78.43 | 75.13 | 78.21 | 76.378 |
Only encoder-decoder convolution block | 83.76 | 77.91 | 83.32 | 81.11 | 82.36 | 81.692 |
MDU-CNN | 82.25 | 79.43 | 83.16 | 81.02 | 85.08 | 82.188 |