Table 3 Comparison of our work with other studies concerning with BPE segmentation with CNN-based models
REF | Approach | Model | Dataset | MRI scanners | DSC | r | Comment |
|---|---|---|---|---|---|---|---|
21 | Segmentaion of FGT from native images, transfer of the mask to subtraction images. Segmentation of BPE based on mean and std of intenisty values | 3D V-Net | 794 patients with unilateral breast cancer (healthy breast was segmented) | 3.0 T: Siemens Verio Phillips Ingenia 1.5 T: GE Signa | Breast 0.91 ± 0.04 FGT 0.85 ± 0.11 | Breast: 0.96 FGT: 0.93 | 3.0 T Siemens and Phillips data in the training and testing set, seprate test set with GE 1.5 T data Evaluation of BPE segmentation not reported |
29 | Segmentation of BPE from subtraction images | 2D U-Net | 38 patients (slices not depicting tumor) | 3.0 T: Siemens Skyra | Overall: 0.76 | - | Only BPE segmentation |
Our work | Segmentation of FGT from native images, independent segmentation of BPE from subtraction images | 2D attention U-Net | 88 patients (slices not depicting tumor) | 1.5 T: Siemens, Sola 3.0 T: Siemens Skyra (two hospitals) | FGT model: Breast 0.950 ± 0.002 FGT 0.820 ± 0.005 (0.864 ± 0.004 wDSC) BPE model: Breast 0.927 ± 0.001 BPE 0.628 ± 0.018 (0.715 ± 0.015 wDSC) | FGT model: Breast 0.999 ± 0.001 FGT% 0.985 ± 0.001 BPE model: Breast 0.992 ± 0.001 BPE% 0.963 ± 0.004 | Data coming from only one scanner used for model training and validation |