nihaomiao / miccai22_adn Goto Github PK

May I ask whether the dice evaluation metrics of the paper is dice per case or global dice?

That is a promising job. Will the complete process of training and testing be released soon?
Thanks!

Refer to some of your suggestions, have some results and questions, would like to discuss with you.
1、May I ask you which normalization method is used in data preprocessing?
I tried two normalization methods, one was global maximum and minimum normalization, and the other was clip to [0,80] , then followed by maximum and minimum normalization. When pretrain with your checkpoint, you can see a big difference in initial results. The result of clip first will be better. Did you clip during normalization? What is the clip range?
2、About the brain tissue segmentation network, why is the output channel 5 instead of 4?
I think background, gray matter, white matter, cerebrospinal fluid, there are four categories.

3、Do you still have the checkpoint of the brain tissue segmentation network(‘’gwm_seg_model‘’)?
4、The following results are my best results at present. There will be some anomalies in the test of some data. I am still investigating the reasons.

First of all, thank you for sharing. I have been reproducing your paper and code recently, and I have some questions to ask you.
(1)The first Network I trained, Transformation Network T, used NVIDIA 3060 for one day, but the result was not very good. May I ask how long have you trained this network?

(2)For the mask in the AISD dataset, is it to use all 1,2,3,5 in the label as labels of the lesion area?
(3)Could you share the model parameters(checkpoint) of the three networks you trained T, D and F, so as to directly test the segmentation effect on the ncct images ?

I'm sorry to bother you again. I downloaded SPM12 and would like to inquire about the specific details of skull dissection and GWM segmentation. Is your GWM segmented using the NII.gz file from CT?

“Yes, as mentioned in our paper, "after skull stripping, we resample all NCCTs to be 1.2 × 1.2 × 5 mm^3 and reshape
them to be 256×256×40 using the Python package nilearn."

“I think it is unreasonable to reshape the image to 40256256 after resampling. You should use the original volume to calculate the relevant metrics, because the asymmetry after resampling cannot represent the asymmetry of the original image.”

we resample all NCCTs to be 1.2 × 1.2 × 5 mm3 and reshape them to be 256 × 256 × 40 using the Python package nilearn. 这里resample然后risize，这样的话resample就失去意义了，我的意思是应该滑动patch，但是因为对称性必须关注全图，这里就自相矛盾了

Could you please provide the checkpoint of the brain tissue segmentation network(‘’gwm_seg_model‘’) ?

Can I add your WeChat account to inquire about the related issues of ADN article? If possible, I am willing to pay you money in return.