Shortcut 3D U-Net powered by Tensorflow 2.0 and the fast preprocessing pipeline with TFRecord for BraTS 2020 challenge
If you don't have data, you can download it from CBICA after registration. Or you can download older BraTS version from Medical Segmentation Decathlon in public.
- Python 3
- tensorflow >= 2.0
- tensorlayer (based on your tensorflow version)
- tensorflow_addons (based on your tensorflow version)
You can modify the following code based on the below descriptions.
# Excerpt code beginning from line 84 in load_data.py
features = {'data': tf.io.FixedLenFeature([], dtype=tf.string),
'label': tf.io.FixedLenFeature([], dtype=tf.string),
'tumor_type': tf.io.FixedLenFeature([], dtype=tf.string),
'name': tf.io.FixedLenFeature([], dtype=tf.string)}
parsed_features = tf.io.parse_single_example(tfr, features)
# The shape of dimension, 4, mean the four modalities provided by MRI images.
data = tf.reshape(tf.io.decode_raw(parsed_features['data'], out_type=tf.float32), self.decode_size+tuple([4]))
# The shape of dimension, 3, mean the three sort of true brain tumors,
# must be the value 0 (background) or 1 (true) in the order of NET (non-enhancing tumor), ED (edema) and ET (enhancing Tumor).
label = tf.reshape(tf.io.decode_raw(parsed_features['label'], out_type=tf.float32), self.decode_size+tuple([3]))
# self.decode_size is (width, height, depth) >>> (240, 240, 155) in default.
# You can change width and height based on your GPU memory size.You need to generate TFRecords with the above format. The variables, tumor_type and name, are redundant and used for recording the brain tumor type and sample name only. If you dont't need it, you can set them as fixed strings when you generate TFRecords.
# Excerpt code beginning from line 74 in load_test_data.py
features = {'data': tf.io.FixedLenFeature([], dtype=tf.string),
'name': tf.io.FixedLenFeature([], dtype=tf.string)}
parsed_features = tf.io.parse_single_example(tfr, features)
data = tf.reshape(tf.io.decode_raw(parsed_features['data'], out_type=tf.float32), self.decode_size+tuple([4]))The format of test TFRecord is similar to load_data.py. The differentiation is that we don't know true labels and brain tumor types.
You can modify the architcture of neural network here and some parameters.
Configure parameters and the path of TFRecord or summaries to get the best results and test model more easily in train.py.
- Ubuntu 20.04
- 12 Intel vCPU
- 1 NVIDIA Tesla V100s 32GB
- 64G RAM
These results provided by leaderboard is computed by CBICA Image Processing Portal online with 125 cases in BraTS 2020 validation dataset.
| Dice_ET | Dice_WT | Dice_TC | Hausdorff95_ET | Hausdorff95_WT | Hausdorff95_TC |
|---|---|---|---|---|---|
| 0.7539 | 0.8962 | 0.82134 | 32.65874 | 5.35762 | 6.57599 |