Table 3:
Values of the performance metrics for tumor segmentation by different models. The shown values correspond to the average and standard deviation (in parenthesis) per fold in five-fold cross-validation. LR represents the used learning rates for training the models.
Model | Training Setting | DSC | JI (IOU) | TPR | FPR | ACC | AUC-ROC |
---|---|---|---|---|---|---|---|
Seg-Net | LR=8·10−4, decreased by 0.5 after 10 epochs until 1·10−4 | 0.889 (±0.011) | 0.811 (±0.015) | 0.877 (±0.019) | 0.088 (±0.014) | 0.977 (±0.002) | 0.957 (±0.004) |
DenseNet-26 | LR=1·10−3, decreased by 0.1 after 10 epochs until 1·10−4 | 0.888 (±0.016) | 0.818 (±0.017) | 0.886 (±0.019) | 0.093 (±0.025) | 0.978 (±0.002) | 0.958 (±0.005) |
PSPNet-ResNet18 | LR=1·10−4, decreased by 0.5 after 10 epochs until 5·10−5, images size of 384×384 pix. | 0.886 (±0.008) | 0.808 (±0.008) | 0.884 (±0.014) | 0.107 (±0.016) | 0.976 (±0.002) | 0.953 (±0.005) |
Ours: U-Net-SA | LR = 1·10−4 | 0.901 (±0.013) | 0.832 (±0.014) | 0.904 (±0.016) | 0.092 (±0.008) | 0.979 (±0.001) | 0.955 (±0.002) |