hilab-git / cople-net Goto Github PK

Can you provide an example of the implementation of the dual branch architecture and DAST (Divergence-Aware Selective Training) as reported in your paper "Learning COVID-19 Pneumonia Lesion Segmentation From Imperfect Annotations via Divergence-Aware Selective Training"?

Dear @taigw ,
Thank you for sharing this helpfull project.
I have occured one problem when reading the DICOM CT volume (.nii.gz) and its labeled image(.nii.gz)。
In the train function of net_run_agent.py, when i run this sentence "labels_prob = self.convert_tensor_type(data['label_prob'])", the error "KeyError: 'label_prob'" appeared.
I have checked the loaded training data, there is a attribute named "label", but not "label_prob".
Could you please help me to solve this problem?
Is there some requirements for the input labeled images? Thanks!

Dear thank you for your work, I used your dataset in my experiments. Could you please provide the source and description of the sample dataset dataset1，dataset2 and dataset3?

Please I need help in training.
I have Nifti images of CT scans. What is the correct order of padding/resizing and random cropping? How to load a 3d volume and then trasnform it for coplenet?

Dear, your work is great, I have some questions and look forward to your reply.
For segment COVID-19 pneumonia lesions from my images, Can you share the code for data pre-processing?
you said that the images have been cropped into the lung region, and the intensity has been normalized into [0, 1] using window width/level of 1500/-650.

Dear @taigw ,

Thanks for sharing the great work.

I'd like to train COPLE-Net with the 70 noisy-label cases.
Following is the config info.

[dataset]
# tensor type (float or double)
tensor_type = float


root_dir  = /path to/part1
train_csv = config/train/all.csv
valid_csv = config/train/valid.csv
test_csv  = config/test/test_all.csv

# modality number
modal_num = 1

# data transforms
# train_transform = [ChannelWiseNormalize, LabelConvert, RandomCrop, LabelToProbability]
# test_transform  = [ChannelWiseNormalize]
train_transform = [Pad, LabelToProbability]
test_transform  = [Pad]

# parameter of Pad
Pad_output_size = [1, 32, 32]
Pad_ceil_mode   = True
Pad_inverse     = True

ChannelWiseNormalize_mean = None
ChannelWiseNormalize_std  = None
ChannelWiseNormalize_channels = [0]
ChannelWiseNormalize_zero_to_random = False
ChannelWiseNormalize_inverse = False

LabelConvert_source_list = [0, 255]
LabelConvert_target_list = [0, 1]
LabelConvert_inverse = False

RandomCrop_output_size = [240, 240]
RandomCrop_foreground_focus = False
RandomCrop_foreground_ratio = None
RandomCrop_mask_label       = None
RandomCrop_inverse     = False

LabelToProbability_class_num = 2
LabelToProbability_inverse   = False
# =====================================================

[network]
# this section gives parameters for network
# In this example, a customized network is used.

[training]
# device name" cuda:n or cpu
device_name = cuda:1

batch_size    = 1
loss_function = noise_robust_dice_loss
noise_robust_dice_loss_p = 1.5
# for optimizers
optimizer     = Adam
learning_rate = 1e-4
momentum      = 0.9
weight_decay  = 1e-5

# for lr schedular (MultiStepLR)
lr_gamma      = 0.5
lr_milestones = [10000]

summary_dir       = model/coplenet
checkpoint_prefix = model/coplenet

# start iter
iter_start = 0
iter_max   = 200000
iter_valid = 1000
iter_save  = 10000

I set batch_size = 1, and my GPU memory is empty (12G available for training).
However, it is wired that out of memory occurred.

RuntimeError: CUDA out of memory. Tried to allocate 184.00 MiB (GPU 1; 11.91 GiB total capacity; 11.29 GiB already allocated; 81.06 MiB free; 11.30 GiB reserved in total by PyTorch)

Would it be possible for you to tell me what's wrong with my settings?

Looking forward to your reply.

Best regards,
Jun

Hi taigw @taigw,

How can I process the image so that its intensity is normalized to [0,1] using window/level of 1500/-650?
Could you please share the code?

Thank you for your help!

Hello,

Can you clarify what is the order in the preprocessing? Do you normalize window/width and to 0, 1 range before or after cropping to the lung region?

I have run your model on some unseen test data and it is not doing well, I did the intensity normalization after cropping.