wuziyi616/cfun

Combining Faster R-CNN and U-net for efficient medical image segmentation

cfun's Issues

About edge_loss

hi, thanks for sharing your code. I have a question about edge loss. When I added the " compute_mrcnn_mask_edge_loss ", the value of loss became "nan" after some iterations. I don't know how to solve this problem, can you help me?

json file format

Can you provide an example with the format required for the json file?

Incorporate into medical toolkits

Thanks for this nice work. If it makes sense, please consider having this technique incorporated into some preexisitng medical segmentation toolkits:

https://github.com/NifTK/NiftyNet
https://github.com/pfjaeger/medicaldetectiontoolkit
https://github.com/perone/medicaltorch

Doing this can increase the audience of your approach. Thanks.

problem about data preprocessing

Hello Ziyi, I recently tried your CFUN code, which is the same MM-WHS data set as the data you used, and also used the test data segmented by U-Net for training, but the final result seems to be very bad, So wanted to ask how did you preprocess the data? For example, the spacing, direction, and origin of the MM-WHS data are different. Have you ever dealt with these?

Thank you very much for your reply

Problems in test function

In function test(model, limit, save, bbox) I've met some problems.

...
label = nib.load(path_label).get_data().copy()
...
    gt_masks[:, :, :, j][label == j + 1] = 1

I think the label here should equal to the label values(like 500, 805, etc.) instead of the class ids. Same problem exists in the function process_mask() which contains code masks[i][mask == i] = 1.

Hope you could reply me soon!

Btw, thank you for sharing these codes!

the ground truth of testing label

Dear ziyi，
Hello! I have seen other people asked this question, but I don't see the answer. The problem is that in your README.md Frequently Asked Questions 1 there is a link 'Update: the ground-truth label for test images are now available [here]', the [here] link can't be downloaded. maybe it needs a sjtu internal account or what‘ s data in the link？（Uhmmm...Seems like you say that you don't have ground truth testing label） If it's exactly the gt testing lable, could you please send it to my email ? [email protected]
Best wishes！

the link of ground-truth label for test images is unavailable here

Thanks for your work, However, the link of the ground-truth label for test images is unavailable. can you update again? thanks

有几个问题想请教一下

1.我在window和Linux上进行训练时，发现mrcnn的各种指标都为0。这是什么原因呢?是不是训练过程中什么变量不是true？
mrcnn_class_loss: 0.00000 - mrcnn_bbox_loss: 0.00000 - mrcnn_mask_loss: 0.00000 - mrcnn_mask_edge_loss: 0.00000
2.我在Linux上进行训练[stage=beginning],epoch=1000结束后，使用其中的某个权重测试发现其IOU都为0[0,0,0,0,0,0]，这是什么原因呢？
3.测试结束后，模型没有生成有效的分割信息。
希望得到您的解答与回复，十分感谢！

which version of skimage should I install？

question about the formula for the feature pyramid (model.py line 330)

Could you please explain detailedly for me about the formula for the feature pyramid (model.py line 330) ??

roi_level = 4 + (1. / 3.) * log2(h * w * d)

You mention that the formula is following and adapting by the paper of the Feature Pyramid Networks paper. I am still getting confused about how you are using this formula instead of using the original one.
Thank you!!

Problem in test

Thanks for your wonderful work！ When I do the test, I met the problem 'UnboundLocalError: local variable 'nms_keep' referenced before assignment', as shown in Figure. I debugged the code and found the parameters like 'pre_nms_class_ids', ‘pre_nms_rois’, and 'pre_nms_scores' are '[]'. I do not know how to solve it. Is this because the parameters I set in the code are wrong? Or because of the test dataset (I used the data of one subject in the training dataset for testing, because I can not open the website (the label for the GT) that you provided)? Have you done special processing for the test dataset? Hope to hear from you as soon as possible. Thanks in advance.

显存爆炸

您是否是没有支持多卡训练呢？ 我单块10G的2080TI每次都显存爆炸。。。 请问有什么办法改进吗？ （实际上我有4块2080TI只有一块正常工作）

IndexError: too many indices for tensor of dimension 1

I use five pictures from the training set as the testing set, when I test the performance of the model, it appears the error"IndexError: too many indices for tensor of dimension 1".
I guess maybe the value for keep_bool and torch.nonzero(keep_bool) is different.
Do you have any idea about it? Thanks!

Errors while running code

nibabel.deprecator.ExpiredDeprecationError: get_data() is deprecated in favor of get_fdata(), which has a more predictable return type. To obtain get_data() behavior going forward, use numpy.asanyarray(img.dataobj).

• deprecated from version: 3.0
• Raises <class 'nibabel.deprecator.ExpiredDeprecationError'> as of version: 5.0

Can you release your pretrained model for testing

Thanks for your nice work！ Can you release your pre-trained model for testing，thanks

Missing edge magnitude computation

Hello, thanks for sharing it. I found one bug. After the line

CFUN/model.py

Line 967 in d684f22

y_true_final = F.conv3d(y_true_[j, :, :, :].unsqueeze(0).unsqueeze(0).cuda().float(), kernel)

I think you should compute the magnitude of the edge for prediction and ground-truth then uses MSE. The above code does not provide edge information. I think it should be

y_true_final  = torch.sqrt(torch.pow(y_true_final[:, 0], 2) + torch.pow(y_true_final[:, 1], 2) + torch.pow(y_true_final[:, 0], 2))
y_pred_final = torch.sqrt(torch.pow(y_pred_final [:, 0], 2) + torch.pow(y_pred_final [:, 1], 2) + torch.pow(y_pred_final [:, 0], 2))

RuntimeError: CUDA out of memory

Thanks for your pretty work.
When I tried to run the program, I found that as the training epoch increased, so did the Memory-Usage, and when I run the 5 epoch,There was an error：
RuntimeError: CUDA out of memory. Tried to allocate 108.00 MiB (GPU 1; 10.91 GiB total capacity; 9.57 GiB already allocated; 32.38 MiB free; 151.
As mentioned in the paper, I use an NVIDIA GeForce GTX 1080Ti GPU, too.
Can you help me with this problem？Thanks very much.

Accuracy of Automated Algorithm

Hi, Dear Author, have you evaluated the accuracy of the automated algorithm for test samples?

In the arXiv CFUN paper, we mentioned in page 7 section 4.1 [that:]
An automated algorithm is designed to generate segmentation labels of 40 originaltest samples...

how to use this model for my own dataset?

i could not get dataset.json file creation step. Can you plesae explain?
Thanks

About the json file

Hello, thank you very much for providing the code, this is a great job, I am a beginner, can you share the json file for my study reference, thank you very much!

pretrained model

Hi

Could you provide pretrained model?

Zhiyuan