Visit our documentation for installation, tutorials and more.
- MRM NeAt segmentation
- Ablation study loss curve
- Pretrained weight
- More examples
- Rater study
- Dataset release
- Contributing to MouseGAN++
- Hyper-parameter
- Limitation
We use 5-fold, 100 epochs, augmentation including resize, shift, scale and rotate for all methods. Details can be found in our paper.
The raw data file and other metrics (e.g., ASD) can be found in Excel table.
| Fold (Dice) | Hipp | SC | Striatum | Tha | Weight file |
|---|---|---|---|---|---|
| Fold 1 | 0.9033 ± 0.0143 | 0.8765 ± 0.0215 | 0.9179 ± 0.0108 | 0.9276 ± 0.0087 | Link (Code: l4oh) |
| Fold 2 | 0.9078 ± 0.0139 | 0.8646 ± 0.0190 | 0.9142 ± 0.0143 | 0.9259 ± 0.0055 | Link (Code: l4oh) |
| Fold 3 | 0.9084 ± 0.0157 | 0.8741 ± 0.0271 | 0.9175 ± 0.0036 | 0.9327 ± 0.0104 | Link (Code: l4oh) |
| Fold 4 | 0.9068 ± 0.0200 | 0.8710 ± 0.0404 | 0.9167 ± 0.0126 | 0.9237 ± 0.0089 | Link (Code: l4oh) |
| Fold 5 | 0.9046 ± 0.0167 | 0.8712 ± 0.0260 | 0.9205 ± 0.0105 | 0.9295 ± 0.0049 | Link (Code: l4oh) |
| Avg (presented in the paper) |
0.9062 ± 0.0146 | 0.8715 ± 0.0250 | 0.9174 ± 0.0100 | 0.9279 ± 0.0077 | - |
We provide tensorboard log files for ablation study experiments. link (Code: 5ygr).
('L_kl' term represents L2 regularization in both MouseGAN and MouseGAN++, actually.)
How to display:
tensorboard --logdir=run1:MouseGAN++,run2:MouseGAN --host localhost --port=6060
# open in the browser: http://localhost:6060/#scalars&_smoothingWeight=0.985
Translation module:
| Method | Dataset | Weight file |
|---|---|---|
| MouseGAN++ | Multi-Modality Dataset | Link (Code: ummb) |
| MouseGAN | Multi-Modality Dataset | Link (Code: ummb) |
| StarGAN-v2 | Multi-Modality Dataset | Link (Code: ummb) |
Segmentation module:
| Method | Dataset | Weight file |
|---|---|---|
| MouseGAN++ | Multi-Modality Dataset (finetune on T1) | Link (Code: ummb) |
| MouseGAN++ | Multi-Modality Dataset (finetune on T2) | Link (Code: ummb) |
| MouseGAN++ | MRM NeAt Dataset | Link (Code: l4oh) |
| Methods | Synthetic image | Modality |
|---|---|---|
| MouseGAN++ |  |
T2*w -> Others |
| MouseGAN++ |  |
T2w -> Others |
| MouseGAN++ |  |
T1w -> Others |
| StarGAN-v2 |  |
- |
Some failed cases:
| Methods | Input image | Real image | Failed Synthetic image | Modality |
|---|---|---|---|---|
| MouseGAN++ |  |
 |
 |
T1w -> T2w |
| MouseGAN++ |  |
 |
 |
T1w -> T2w |
| StarGAN-v2 |  |
 |
 |
T1w -> T2w |
| StarGAN-v2 |  |
 |
 |
T1w -> T2w |
We have invited three experts to score the synthetic images from 12 test mice with regard to the delineation of related anatomy on a 10-point scale as follows: excellent diagnostic quality (8-10), good diagnostic quality (6-8), fair diagnostic quality (4-6), poor diagnostic quality (2-4), and non-diagnostic (0-2). Please refer to the following table.
| Real image (label) | MouseGAN++ | MouseGAN | CycleGAN | SynSeg | UNIT | MUNIT | StarGAN-v2 | |
|---|---|---|---|---|---|---|---|---|
| Score (mean ± std) |
9.13 ± 0.72 | 7.56 ± 0.84 | 7.25 ± 0.84 | 4.58 ± 1.23 | 4.67 ± 1.15 | 4.36 ± 1.13 | 4.47 ± 1.50 | 4.67 ± 1.45 |
Please send emails to Dr. Xiao-Yong Zhang for the download links.
We are happy about any contributions! (MRI data / trained weight / plug-in function code)
MouseGAN++ follows the open-access paradigm, allowing users to save their updated models and share their weights for use by the neuroimaging community.
Besides, the accumulation of additional imaging data will further improve the performance of MouseGAN++ and support the exploration of complex neuroimaging research.
How to tune these hyper-parameters: The hyper-parameters matter and are task-dependent. They are not carefully selected yet. Despite this, the selection of hyper-parameters reported in our paper works well on mouse brain datasets.
In addition, we conducted an ablation study on contrastive loss hyper-parameters. We discovered that λ=1 works well on our task and maintains balance between other losses during the training procedure.
The quantitative results are provided as follows:
| T1 -> T2 | T2 -> T1 | |
|---|---|---|
| Results |  |
 |
where λ=0 is the baseline method (previous MouseGAN).
One limitation of our work is the dilemma caused by domain shifts, since the image quality and contrast from various centers may differ greatly. An appealing and promising solution is to convert our pretext task to cross-center image translation so that the learned center-agnostic features in the content space would alleviate the segmentation performance degradation.