zaccharieramzi
commited on
Commit
•
bb78876
1
Parent(s):
32474e6
added model card and model weights
Browse files- README.md +86 -0
- model_weights.h5 +3 -0
README.md
ADDED
@@ -0,0 +1,86 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# NCPDNet-3D
|
2 |
+
---
|
3 |
+
tags:
|
4 |
+
- TensorFlow
|
5 |
+
- MRI reconstruction
|
6 |
+
- MRI
|
7 |
+
datasets:
|
8 |
+
- fastMRI
|
9 |
+
---
|
10 |
+
|
11 |
+
This is a non-Cartesian 3D MRI reconstruction model for radial trajectories at acceleration factor 4.
|
12 |
+
The model uses 10 iterations and a small vanilla CNN.
|
13 |
+
|
14 |
+
## Model description
|
15 |
+
For more details, see https://hal.inria.fr/hal-03188997.
|
16 |
+
This section is WIP.
|
17 |
+
|
18 |
+
## Intended uses and limitations
|
19 |
+
This model can be used to reconstruct 3D brain data obtained retrospectively from magnitude scanners of the OASIS database at acceleration factor 4 in a fully radial acquisition setting.
|
20 |
+
|
21 |
+
## How to use
|
22 |
+
This model can be loaded using the following repo: https://github.com/zaccharieramzi/fastmri-reproducible-benchmark.
|
23 |
+
After cloning the repo, `git clone https://github.com/zaccharieramzi/fastmri-reproducible-benchmark`, you can install the package via `pip install fastmri-reproducible-benchmark`.
|
24 |
+
The framework is TensorFlow.
|
25 |
+
|
26 |
+
You can initialize and load the model weights as follows:
|
27 |
+
```python
|
28 |
+
import tensorflow as tf
|
29 |
+
|
30 |
+
from fastmri_recon.models.subclassed_models.ncpdnet import NCPDNet
|
31 |
+
|
32 |
+
|
33 |
+
model = NCPDNet(
|
34 |
+
three_d=True,
|
35 |
+
n_iter=6,
|
36 |
+
n_filters=16,
|
37 |
+
im_size=(176, 256, 256),
|
38 |
+
dcomp=True,
|
39 |
+
fastmri=False,
|
40 |
+
)
|
41 |
+
kspace_shape = 1
|
42 |
+
inputs = [
|
43 |
+
tf.zeros([1, 1, kspace_shape, 1], dtype=tf.complex64),
|
44 |
+
tf.zeros([1, 3, kspace_shape], dtype=tf.float32),
|
45 |
+
(tf.constant([(176, 256, 256)]), tf.ones([1, kspace_shape], dtype=tf.float32)),
|
46 |
+
]
|
47 |
+
model(inputs)
|
48 |
+
model.load_weights('model_weights.h5')
|
49 |
+
```
|
50 |
+
|
51 |
+
Using the model is then as simple as:
|
52 |
+
```python
|
53 |
+
model([
|
54 |
+
kspace, # shape: [n_batch, 1, n_kspace_samples, 1]
|
55 |
+
traj, # shape: [n_batch, 1, 3, n_kspace_samples]
|
56 |
+
(
|
57 |
+
output_shape, # shape: [n_batch, 3]
|
58 |
+
dcomp, # shape: [n_batch, n_kspace_samples]
|
59 |
+
)
|
60 |
+
])
|
61 |
+
```
|
62 |
+
|
63 |
+
## Limitations and bias
|
64 |
+
The limitations and bias of this model have not been properly investigated.
|
65 |
+
|
66 |
+
## Training data
|
67 |
+
This model was trained using the [OASIS dataset](https://www.oasis-brains.org/).
|
68 |
+
|
69 |
+
## Training procedure
|
70 |
+
The training procedure is described in https://hal.inria.fr/hal-03188997.
|
71 |
+
This section is WIP.
|
72 |
+
|
73 |
+
## Evaluation results
|
74 |
+
On the OASIS validation dataset:
|
75 |
+
- PSNR: 33.76
|
76 |
+
|
77 |
+
## Bibtex entry
|
78 |
+
```
|
79 |
+
@unpublished{ramzi:hal-03188997,
|
80 |
+
TITLE = {{NC-PDNet: a Density-Compensated Unrolled Network for 2D and 3D non-Cartesian MRI Reconstruction}},
|
81 |
+
AUTHOR = {Ramzi, Zaccharie and G R, Chaithya and Starck, Jean-Luc and Ciuciu, Philippe},
|
82 |
+
YEAR = {2021},
|
83 |
+
MONTH = Sep,
|
84 |
+
}
|
85 |
+
```
|
86 |
+
|
model_weights.h5
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:1b15e4c3670dc5b08900f0aaa68339b130468ce08c499d5ccb223f962392bcff
|
3 |
+
size 328264
|