Table 6.
Performance comparison of segmentation models built upon images normalized by the robust deviation and the standard deviation (the sizes of patches used to train FCNNs were 33*33*3 and 65*65*3 respectively, n = 5, and the number of patches used to train FCNNs was 5000*5*20).
Dataset | Deviation | Methods | Dice | PPV | Sensitivity | ||||||
---|---|---|---|---|---|---|---|---|---|---|---|
|
|
|
|||||||||
Complete | Core | Enhancing | Complete | Core | Enhancing | Complete | Core | Enhancing | |||
Challenge | standard | FCNNs | 0.73 | 0.69 | 0.67 | 0.61 | 0.60 | 0.61 | 0.94 | 0.86 | 0.75 |
FCNN + CRF | 0.84 | 0.80 | 0.71 | 0.87 | 0.79 | 0.68 | 0.82 | 0.82 | 0.75 | ||
FCNN + CRF + post-process | 0.86 | 0.83 | 0.76 | 0.93 | 0.86 | 0.80 | 0.81 | 0.81 | 0.74 | ||
robust | FCNNs | 0.74 | 0.72 | 0.67 | 0.62 | 0.63 | 0.60 | 0.94 | 0.86 | 0.77 | |
FCNN + CRF | 0.85 | 0.80 | 0.70 | 0.87 | 0.80 | 0.63 | 0.84 | 0.81 | 0.80 | ||
FCNN + CRF + post-process | 0.87 | 0.83 | 0.76 | 0.92 | 0.87 | 0.77 | 0.83 | 0.81 | 0.77 | ||
Learderboard | standard | FCNNs | 0.69 | 0.60 | 0.54 | 0.57 | 0.55 | 0.50 | 0.97 | 0.75 | 0.67 |
FCNN + CRF | 0.83 | 0.66 | 0.58 | 0.85 | 0.71 | 0.56 | 0.85 | 0.70 | 0.67 | ||
FCNN + CRF + post-process | 0.86 | 0.73 | 0.61 | 0.89 | 0.75 | 0.66 | 0.84 | 0.78 | 0.66 | ||
robust | FCNNs | 0.70 | 0.61 | 0.54 | 0.58 | 0.57 | 0.49 | 0.96 | 0.74 | 0.67 | |
FCNN + CRF | 0.83 | 0.66 | 0.57 | 0.85 | 0.71 | 0.50 | 0.85 | 0.69 | 0.71 | ||
FCNN + CRF + post-process | 0.86 | 0.73 | 0.62 | 0.89 | 0.76 | 0.64 | 0.84 | 0.78 | 0.68 |