Table 4 Pixel-level and object-level dice coefficient for crypt segmentation of synthetic images of various grades at 20 × magnification. Crypts were segmented using a thresholded probability map method [3]. Results are shown when the method was trained and tested on the synthetic and on real data. The reported figures are the average ± standard deviation
From: A model of the spatial tumour heterogeneity in colorectal adenocarcinoma tissue
Training data | Test | Grade | Dice-Pixel | Dice-Object |
|---|---|---|---|---|
Synthetic | Synthetic | Healthy | 0.96 ± 0.003 | 0.91 ± 0.03 |
|  |  | Well | 0.94 ± 0.005 | 0.90 ± 0.03 |
|  |  | Moderately | 0.91 ± 0.02 | 0.90 ± 0.03 |
|  |  | Poorly | 0.65 ± 0.15 | 0.65 ± 0.13 |
Real | Synthetic | Healthy | 0.87 ± 0.01 | 0.85 ± 0.02 |
|  |  | Well | 0.89 ± 0.01 | 0.84 ± 0.03 |
|  |  | Moderately | 0.88 ± 0.11 | 0.52 ± 0.11 |
|  |  | Poorly | 0.59 ± 0.16 | 0.36 ± 0.11 |
Synthetic | Real | Benign | 0.69 ± 0.11 | 0.53 ± 0.13 |
|  |  | Moderately | 0.58 ± 0.16 | 0.43 ± 0.13 |
|  |  | Poorly | 0.60 ± 0.17 | 0.44 ± 0.17 |