zaccharieramzi
commited on
Commit
•
2c83c8e
1
Parent(s):
65d4b01
added model card and model weights
Browse files- README.md +92 -0
- model_weights.h5 +3 -0
README.md
ADDED
@@ -0,0 +1,92 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# UPDNet-knee-af8
|
2 |
+
---
|
3 |
+
tags:
|
4 |
+
- TensorFlow
|
5 |
+
- MRI reconstruction
|
6 |
+
- MRI
|
7 |
+
datasets:
|
8 |
+
- fastMRI
|
9 |
+
---
|
10 |
+
|
11 |
+
This model was used to achieve the 9th highest submission in terms of PSNR on the fastMRI dataset (see https://fastmri.org/leaderboards/) (0.2dB behind the 2nd submission).
|
12 |
+
It is a base model for acceleration factor 8.
|
13 |
+
The model uses 25 iterations and a medium-ca-prelu U-net, and a medium sensitivity maps refiner.
|
14 |
+
|
15 |
+
## Model description
|
16 |
+
For more details, see https://arxiv.org/abs/2010.07290.
|
17 |
+
This section is WIP.
|
18 |
+
|
19 |
+
## Intended uses and limitations
|
20 |
+
This model can be used to reconstruct knee data from Siemens scanner at acceleration factor 8.
|
21 |
+
|
22 |
+
## How to use
|
23 |
+
This model can be loaded using the following repo: https://github.com/zaccharieramzi/fastmri-reproducible-benchmark.
|
24 |
+
After cloning the repo, `git clone https://github.com/zaccharieramzi/fastmri-reproducible-benchmark`, you can install the package via `pip install fastmri-reproducible-benchmark`.
|
25 |
+
The framework is TensorFlow.
|
26 |
+
|
27 |
+
You can initialize and load the model weights as follows:
|
28 |
+
```python
|
29 |
+
import tensorflow as tf
|
30 |
+
|
31 |
+
from fastmri_recon.models.subclassed_models.updnet import UPDNet
|
32 |
+
|
33 |
+
|
34 |
+
model = UPDNet(
|
35 |
+
multicoil=True,
|
36 |
+
n_dual=1,
|
37 |
+
primal_only=True,
|
38 |
+
n_layers=4,
|
39 |
+
n_iter=25,
|
40 |
+
channel_attention_kwargs={'dense': True},
|
41 |
+
refine_smaps=True,
|
42 |
+
non_linearity='prelu',
|
43 |
+
layers_n_channels=[16 * 2**i for i in range(4)],
|
44 |
+
)
|
45 |
+
kspace_size = [1, 1, 320, 320]
|
46 |
+
inputs = [
|
47 |
+
tf.zeros(kspace_size + [1], dtype=tf.complex64), # kspace
|
48 |
+
tf.zeros(kspace_size, dtype=tf.complex64), # mask
|
49 |
+
tf.zeros(kspace_size, dtype=tf.complex64), # smaps
|
50 |
+
]
|
51 |
+
model(inputs)
|
52 |
+
model.load_weights('UPDNet-knee-af8/model_weights.h5')
|
53 |
+
```
|
54 |
+
|
55 |
+
Using the model is then as simple as:
|
56 |
+
```python
|
57 |
+
model([
|
58 |
+
kspace, # shape: [n_slices, n_coils, n_rows, n_cols, 1]
|
59 |
+
mask, # shape: [n_slices, n_coils, n_rows, n_cols]
|
60 |
+
smaps, # shape: [n_slices, n_coils, n_rows, n_cols]
|
61 |
+
])
|
62 |
+
```
|
63 |
+
|
64 |
+
## Limitations and bias
|
65 |
+
The limitations and bias of this model have not been properly investigated.
|
66 |
+
|
67 |
+
## Training data
|
68 |
+
This model was trained using the [fastMRI dataset](https://fastmri.org/dataset/).
|
69 |
+
|
70 |
+
## Training procedure
|
71 |
+
The training procedure is described in https://arxiv.org/abs/2010.07290.
|
72 |
+
This section is WIP.
|
73 |
+
|
74 |
+
## Evaluation results
|
75 |
+
No evaluation available outside the one from the fastMRI leaderboard (id: `updnet_v3`).
|
76 |
+
|
77 |
+
|
78 |
+
## Bibtex entry
|
79 |
+
```
|
80 |
+
@inproceedings{Ramzi2020d,
|
81 |
+
archivePrefix = {arXiv},
|
82 |
+
arxivId = {2010.07290},
|
83 |
+
author = {Ramzi, Zaccharie and Ciuciu, Philippe and Starck, Jean-Luc},
|
84 |
+
booktitle = {ISMRM},
|
85 |
+
eprint = {2010.07290},
|
86 |
+
pages = {1--4},
|
87 |
+
title = {{XPDNet for MRI Reconstruction: an application to the 2020 fastMRI challenge}},
|
88 |
+
url = {http://arxiv.org/abs/2010.07290},
|
89 |
+
year = {2021}
|
90 |
+
}
|
91 |
+
```
|
92 |
+
|
model_weights.h5
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:8ae0201a7b9e9013bf748191d0ed9eb6bcca08d614e0dfb70241cd26abe91ad2
|
3 |
+
size 50442672
|