Datasets have been uploaded to google drive. You can access these via the following URL:
Original TrainDatasets: https://drive.google.com/file/d/1WQdSuA_mj-d57oOnQ8VYpTzTFAnfGc2w/view?usp=sharing
TestSets: https://drive.google.com/file/d/1NpPXQ6UYkGySZMC9Q34b4fKYXhxHPiEF/view?usp=sharing
Note that the above datas are 2D-images, you need to compose them to 3D shape.
This part is coded by matlab (python, Yes!) and the specific workflow had been elaborated in the paper-"CT-image Super Resolution Using 3D Convolutional Neural Network-Section 3.2”. . (Due to the publication consideration, we haven't published the latest code of how to make datas, if you are editors or reviewers, please concat me with 1556905690@qq.com)
The training data is generated through 5 sets of CT images, summed up to 2000 single 2D images totally. So this is so big training data library. Finally, We have generated 32.5G of data, it is inconvenient to upload to Github. If your would like to using our data to train, please concat me with 1556905690@qq.com or issue in this Repository.
It is easy to train our network,running with specified parameters, and the following is a helper of parameters:
python main.py
--batchSize", type=int, default=64, help="Training batch size"
--nEpochs", type=int, default=100, help="Number of epochs to train for"
-lr", type=float, default=0.1, help="Learning Rate. Default=0.1"
--step", type=int, default=10, help="Sets the learning rate to the initial LR decayed by momentum every n epochs, Default: n=10"
--cuda", type=int,default=1, help="Use cuda?"
--resume", default="", type=str, help="Path to checkpoint (default: none)"
--start-epoch", default=1, type=int, help="Manual epoch number (useful on restarts)"
--clip", type=float, default=0.4, help="Clipping Gradients. Default=0.4")
--threads", type=int, default=1, help="Number of threads for data loader to use, Default: 1"
--momentum", default=0.9, type=float, help="Momentum, Default: 0.9"
--weight-decay", "--wd", default=1e-4, type=float, help="Weight decay, Default: 1e-4"
--pretrained', default='', type=str, help='path to pretrained model (default: none)'
--train_path',type=str,default="train_data/3dtrain25all.h5",help='Path to train dataset'
--memo', default= 'L_', type=str, help='prefix of logger '
3dsrcnn.py This program is a structure for building a network. ``
Because our validation dataset is very complex which contains 5 sets of CT images, each set contains x2, x3, x4 scale. What's more, we need to calculate PSNR of every Epoch. We provided automatically executing script.
Note that you must specify path to low resoluiton images(inter_path)
original HR path (ori_path)
bash auto-execute.sh
if you want compare multi-scale or model of specified epoch, you could run
bash auto-excute=ulti_testset.sh
For example, the generated log is follwing:
x2, x3, x4 scale PSNR of model/0310-2114_model/model_epoch_35.pkl
-----------------
2019/3/ 23 Tue 07:21:31 CST
33.37136136045343
30.771517328760474
35.299094237934455
39.6395555291764
45.80099485177503
------------------
2019/3/ 21 Tue 08:04:13 CST
29.78923091882109
26.636256131118948
32.5371038561331
37.378719766068414
44.580486528475156
-----------------
2019/ 3 21 Tue 08:46:30 CST
26.629647924465274
23.407602537288376
31.092819773019947
35.71189104481404
42.932984128665936
-----------------
Note that: Each scale have five sets of CT images, so it will ouput five values, the acerage PSNR. You can calculate the average PSNR mannually. And the generated results will store in output.txt(you also can specified the fileName). Note that, If you want to use
auto-execute.sh, you must promise established file-tree format!
ori_path='30x30_2'#HR_path
inter_path='30x30_2'#LR_path(under feeding into network)
model_path='model_12layers_25input_3kernel_multi/model_epoch_20.pkl'#specify model path
saved_prefix='30x30_'#prefix file path to save generated images
format='bmp'#reading images format
- Why do we restore with small blocks? If you feed size of 400x400x400 blocks in to network, I think that your computer would explode because of the limitaitons of GPU memory(Unless you have extra large memory). So we need to reconstruct by sub-blocks and compose those by order.