Hello, I read the paper carefully.

I was interested in the research and tried the code you provided.

However, after running the code, I found that the performance was significantly lower than what was reported in the paper, so I decided to write this post to ask questions about the code.

Could you please let me know if I missed anything?

First, let me explain how I reproduced the code.

1. The main code that runs the training is python ./code/LA_M3HL_train.py

2. The code missing from the code for implementing this code was taken from the BCP paper.
   This is because the overall framework of the uploaded code and the BCP code are very similar.
   (code that should be in the utils file, etc.)

3. I modified that code in the main code.

I implemented it by adding the following code with reference to the code in ACDC_M3HL_train.py.

4. It may be a problem with the current computer specifications, so to confirm, I ran BCP, which was included in the comparison group of the paper, and the performance was similar.

(The results above are the BCP presented in your paper, and the results below are the BCP I ran. )

5. Despite this, the performance claimed in the paper is 91% of DICE, while the performance achieved when running the current code is 89%.

If you see this post, I would appreciate it if you could leave a reply to this question.