search

xmed-lab / DHC

MICCAI 2023: DHC: Dual-debiased Heterogeneous Co-training Framework for Class-imbalanced Semi-supervised Medical Image Segmentation
MIT License
51 stars 3 forks source link

Questions Regarding Spynase Dataset Split and Result Discrepancies #4

Closed lzhang30 closed 10 months ago

lzhang30 commented 10 months ago

Hello,

First of all, I'd like to extend my gratitude for your open-source project. It has significantly aided my research. I am writing to seek clarification regarding certain aspects of the Synapse dataset segmentation.

To begin with, I have a preliminary question regarding the order of classes: Is the order of classes shown in the logs the same as the one presented in the tables in your paper? If they are the same, I have observed a notable discrepancy in the data of the fourth column. If they differ, could you please specify the order of classes as they appear in the log output?

Moving on to more specific queries:

  1. Concerning the Synapse dataset used in your project, is it identical to the RawData.zip found in the BTCV dataset? I encountered a hurdle while executing your preprocess.py script, as I couldn't locate an imagesTr folder in the raw data. Subsequently, I downloaded the data from the link you provided and modified the code in evaluate_Ntimes.py as shown in the attached image. However, this led to a size mismatch due to the patch-size settings in your preprocessing script (as seen in the attached image). The expected resize shape through this code should be (80,160,160), causing a size mismatch error (details in the attached image), since the size of the prediction inferred from test.py may be in a similar shape as the image volume (72,144,144).

Following this, I conducted two experiments with different shapes: one with the raw data from RawData.zip resized to 80,160,160, and another used the data downloaded from one-drive in the shape of 72,144,144. The results of these experiments are attached. While the results for other columns were quite close, there was a significant variance in the results of the fourth column, even when tested on datasets of 40% and 20% using the split you provided.

I am keen to understand these discrepancies and would greatly appreciate your insights or any guidance you could provide. Thanks again for your wonderful open-source project!

mismatch error:

The patch-size set in preprocess was
image, and the resize-shape through this code should be like (80,160,160), cause the error of size mis-match image because the size of the pred inffered from test.py may in the same shape like the image volume as (72,144,144).

Result: (in 20% synapse dataset) in 80,160,160: image in 72,144,144 image Futher more, I trained fully-Vnet by usingpython ./code/train_fully.py, and I got the result like this: image It is a little far form the full-VNet's result in the paper.

McGregorWwww commented 10 months ago

Thanks for your attention and pointing out the issue. Actually, it is the code version issue, since when we conducted the AMOS experiment, we changed a lot of the code. Now we update several scripts and the pre-processed data. Please download them again and sorry for the inconvenience.

As for the fully supervised performance, maybe you can change the epoch to 800, thus the performance will be more stable.