Introduction
Materials and methods
Data acquisition
Network architectures and optimization
-
in the GAN-based approach, where usually the segmentation loss relies on the source label map (since the transformation between source and target is learnt), we added true target labels. Hence, the segmentation loss was alternatively obtained by source-to-target label maps or real target label maps, depending on the input case;
-
in the VAE-based approach, we used a similar technique. Here, the segmentation loss is originally just for the source image, we added one for the target, and we conditioned the reconstruction of the target image not solely on the predicted label map, but on the ground truth label map, when available.
Evaluation setup
Results
Method | SUP | LV | LVM | RV | LA | RA | AO |
---|---|---|---|---|---|---|---|
GT | 99.64 ± 33.14 | 115.51 ± 26.63 | 117.32 ± 31.92 | 57.45 ± 20.31 | 70.18 ± 21.97 | 35.90 ± 9.23 | |
VAE | 0% | 35.6 ± 17.7 (40.5) | 37.9 ± 28.6 (47.3) | 63.3 ± 25.9 (68.2) | 32.2 ± 15.5 (35.7) | 26.9 ± 12.4 (29.6) | 13.7 ± 9.1 (16.4) |
10% | 12.6 ± 10.6 (16.4) | 31.4 ± 16.1 (35.2) | 39.7 ± 19.9 (44.3) | 20.2 ± 16.4 (25.9) | 19.4 ± 10.7 (22.1) | 8.6 ± 9.8 (13.0) | |
20% | 6.3 ± 9.2 (11.1) | 28.2 ± 13.1 (31.1) | 33.8 ± 20.0 (39.2) | 13.9 ± 14.5 (20.0) | 18.3 ± 9.5 (20.6) | 9.3 ± 5.4 (10.7) | |
30% | 3.4 ± 8.6 (9.2) | 26.5 ± 11.5 (28.8) | 24.5 ± 15.7 (29.0) | 12.2 ± 12.9 (17.7) | 16.3 ± 11.3 (19.8) | 7.5 ± 5.4 (9.2) | |
GAN | 0% | 12.7 ± 12.4 (17.6) | 3.9 ± 14.1 (14.5) | 20.8 ± 13.1 (24.5) | 2.4 ± 16.3 (16.2) | − 0.8 ± 9.2 (9.2) | 1.8 ± 4.5 (4.8) |
10% | 2.3 ± 7.2 (7.5) | 2.1 ± 8.8 (8.9) | 7.4 ± 11.2 (13.4) | 4.7 ± 13.7 (14.3) | 1.2 ± 10.4 (10.3) | 1.0 ± 5.2 (5.2) | |
20% | 0.0 ± 7.6 (7.5) | − 1.4 ± 12.4 (12.3) | 3.6 ± 11.2 (11.6) | 2.0 ± 13.8 (13.8) | − 0.4 ± 9.7 (9.6) | 0.6 ± 5.4 (5.4) | |
30% | − 1.8 ± 8.1 (8.3) | 0.6 ± 9.5 (9.4) | 2.2 ± 10.3 (10.5) | 1.9 ± 10.9 (11.0) | 0.6 ± 9.6 (9.6) | 0.0 ± 5.4 (5.4) | |
UNet | 10% | 47.2 ± 75.1 (87.9) | − 5.1 ± 27.4 (27.6) | − 9.8 ± 25.6 (27.1) | 7.6 ± 42.5 (42.7) | 7.4 ± 32.9 (33.3) | − 5.9 ± 9.3 (10.9) |
20% | 97.2 ± 73.2 (121.2) | 0.3 ± 14.9 (14.7) | − 9.4 ± 18.1 (20.3) | − 1.7 ± 19.7 (19.5) | 0.4 ± 17.3 (17.1) | 1.1 ± 11.1 (11.1) | |
30% | 15.6 ± 46.6 (48.7) | 6.7 ± 19.9 (20.8) | − 2.0 ± 18.5 (18.4) | 3.1 ± 27.2 (27.0) | − 2.6 ± 15.5 (15.5) | − 0.3 ± 6.4 (6.4) |