DICE

[H-CODE6]

Hello, your code has been reproduced recently. We can't know the Dice coefficient after the test. How did you get the coefficient

[BruceResearch]

Hi, I added a new folder name "evaluation" for you to calculate the Dice score for the validation data. However, this equation for my Dice score may not be the same as that for the BraTS Challenge. If https://www.synapse.org/#!Synapse:syn25829067/wiki/611501 still accepts validation submission, you might want to download the official validation data and upload your segmentation for an official Dice score to evaluate your trained model.

[H-CODE6]

Thank you. Can you provide a trained model epoch last. PTH so that I can reproduce your effect.

[BruceResearch]

So you would like me to upload my trained final model?

[H-CODE6]

Yes, thank you for your reply.

[BruceResearch]

https://drive.google.com/file/d/13Vom6579mXtt0M3AnXt9pWIlhIDckTMM/view?usp=sharing I have uploaded the trained model that I submitted for the testing phase of BraTS 2021 on Google Drive. Please unzip it for the pth file.

[H-CODE6]

OK, thank you very much for your reply.

[BruceResearch]

No problem. Note that this model is trained with all the training data.

[H-CODE6]

Have you trained all the data of brats2021? What test set and verification set do you use for testing?

[BruceResearch]

Validation data can be downloaded from https://www.synapse.org/#!Synapse:syn25829067/wiki/611501 However, the ground-truth segmentation is unavailable for us. Thus, I need to submit the segmentation images for the website to validate the score for us. That was how I did last year, so I did not need to do the training/validation split. There is a detailed introduction on that website. However, I am not sure if they are still accepting validation requests.

[H-CODE6]

OK, thank you for replying to my question. Thank you very much.

[BruceResearch]

You are welcome.

[H-CODE6]

Hello, author. Do you have a pre training model in the resume when you first started training? Or do you just train to get the model epoch Last, finally test.

[BruceResearch]

https://drive.google.com/file/d/1YAVylEuPz1DDZOTUPdS2YGGYquuOUtYs/view?usp=sharing

Uploaded another model with 1600 epochs, which is the one with the fewest epochs I can possibly find.

[BruceResearch]

Yes, I continually saved my models and validate their performances while training.

[H-CODE6]

OK, if I want to train my model, is it running train Py, nothing is filled in the resume, but I find a line of re-training prompt when running, and then start training. Does this sentence have any impact? Or the normal operation process?

[BruceResearch]

line 102 determines the save frequency, the default is to save the model for every 50 epochs.

Yeah I read through the code, if u provide a trained model for resuming, it will print "Successfully loading checkpoint {} and training from epoch: {}" If you dont provide and leave resume empty, it will print 're-training!!!'

Sorry for the confusion as it shall not say 're-training!!!' when u are training from epoch 0.

[H-CODE6]

Yes, I want to train my own model. At the beginning, I don't fill in anything in the resume, and finally get my own training model epoch Last, I wonder if my understanding is correct?

[BruceResearch]

yes, just leave the resume blank

[H-CODE6]

Thank you for your answer. Thank you very much. I ran your evaluation code today and found that only WT, TC, mean dice, et and mean hausdorff95 (mm) in your paper can not be evaluated. I wonder if it will take your valuable time to update the evaluation code so that we can better understand the paper and compare the results.

[BruceResearch]

Sorry, I never wrote or adapted any code for calculating hausdorff95 (mm). All the results that I obtained are from the validation from the official BraTS host. Please look up other repos to adopt the function of calculating hausdorff95 (mm) for BraTS data.

[H-CODE6]

OK, thank you very much for your reply. Since I haven't conducted the evaluation on the official brats host again, is it convenient for you to tell me the evaluation process?

[BruceResearch]

There is a very detailed instruction on https://www.synapse.org/#!Synapse:syn25829067/wiki/611501 You just need to uplead the segmentation files following a certain format, which my code already handles.

[H-CODE6]

OK, thank you very much for your answer

[H-CODE6]

Hello, data postprocessing cannot run successfully after the test. Can you tell me which parameters to change in the code? Or download those files?

[BruceResearch]

The performance of the method of connected_components.py is not ideal after experiments, and in my paper we use majority voting only. Thus, please just run ensemble_by_majority_voting.py as your post process procedure. You need to specify test_path (main folder where stores segmentation outputs from each model) and model_list(specific folder names of each model's outputs). When running ensemble_by_majority_voting.py, specify the method, which is" python ensemble_by_majority_voting.py --m 0"

[H-CODE6]

Excuse me, model list is the specific folder name of each model output you refer to on the list? I only generated a model on the checkpoint epoch last. pth ，and other files, as well as the above result path，pred path，prob_ Does path need to be modified?

[BruceResearch]

You can not do the ensemble with only one model. You have to generate several sets of outputs from different models to do the ensemble.

[H-CODE6]

Well, thank you very much.

[H-CODE6]

Hello, what is the best effect of your final experimental results? What are the dice and other coefficients respectively? I want to compare it with the results I reproduced.

[BruceResearch]

Please refer to Table 1 in the paper : https://arxiv.org/pdf/2109.12271.pdf

[H-CODE6]

Yes, I have read your paper completely and am interested in the coefficients of Table 2 and table 3. How do you realize it?

[BruceResearch]

Those summary statistics are provided and calculated by the BraTS official validation result.

[BruceResearch]

We still mainly look at the mean result. My table 2 and table 3 provide other summary statistics, which only serve as a supplementary perspective of evaluating the performance.

[H-CODE6]

OK, thank you very much for your answer

[H-CODE6]

Hello, I just used Dual_focal_loss in the loss function， The performance test of the loss function shows that the image background of the final test output is blue, and the test result is not very normal. Do you want to modify the code when using other loss functions.

[H-CODE6]

Hello, author. I've been studying your code recently, at bitrunet.py Some of are not very clear. Can you comment on the specific functions in front of each function and class? So that our readers can better understand the paper.

[BruceResearch]

Hi, Sorry I have been very busy these days. I have added some comments for each class to illustrate their roles. The main idea is just that I define the encoder class and decoder class separately, and finally, combine them in one def.

[H-CODE6]

OK, thank you very much for your reply to me in your busy schedule. I would like to express my deep gratitude to you again.

[H-CODE6]

Hello, after using your evaluation file, it is found that ET performance cannot be evaluated. Can you update the code to facilitate our evaluation? We look forward to your reply.

[BruceResearch]

I see where the problem is. It should be fixed now. Please let me know if it works.

[H-CODE6]

OK, thank you very much. I'll give you feedback immediately after my evaluation. Thank you again

[BruceResearch]

Just updated again, found a typo

[H-CODE6]

The following error occurred

Traceback (most recent call last): File "evaluation.py", line 90, in dice = dice_of_brats_data_set(gt_names, seg_names, type_idx) File "evaluation.py", line 55, in dice_of_brats_data_set s_volume[s_volume == 2 or s_volume ==1] = 0 ValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all()

[BruceResearch]

what about now

[H-CODE6]

It can be run. At that time, the final result of the same data WT and TC is very different from that of the code that has not been updated before

[H-CODE6]

now： tissue type ET dice mean [0.83201768] dice std [0.23055119] tissue type WT dice mean [0.80649212] dice std [0.1666985] tissue type TC dice mean [0.71619503] dice std [0.28385196]

old： tissue type whole dice mean [0.90952711] dice std [0.10472148] tissue type core dice mean [0.88381648] dice std [0.18333644] tissue type all tissue label [1, 2, 3, 4] dice mean [0.71619503 0.80649212 1. 0.83201768] dice std [0.28385196 0.1666985 0. 0.23055119]

[BruceResearch]

That is right, the old code is totally wrong for the tumor label since I actually did not use them at all. Based on the info on BraTS website, "Annotations comprise the GD-enhancing tumor (ET — label 4), the peritumoral edematous/invaded tissue (ED — label 2), and the necrotic tumor core (NCR — label 1)", I modified the code.

[BruceResearch]

I see, it can still be changed, give me a minute.

[BruceResearch]

updated

[H-CODE6]

No problem. Can you tell me why