About TransMIL TCGA Test Accuracy 85.24%?

[lihua8848]

Thank you for your interesting work, but I have questions about the results of TransMIL's comparative experiment in TCGA in the article. The data set is the data set of the DSMILused in your article, and the training set and test set are the same as in the article.

The official code of TransMIL is here, the parameters used are also official parameters, and the results I ran were 93.3% accurate.

How did 85.24% of the article come from

In addition, I also noticed the article SSLP: Spatial Guided Self-supervised Learning on Pathological Images of your team MICCAI2021 a few years ago, which said that the code would be made public, but it has not been made public at present

Can you explain that? Thank you！

[lihua8848]

The training and testing dataset is here in DSMIL official code, and you should split it to training and testing dataset like TransMIL format.

This is the csv we made tcga_NSCLC.CSV

[HHHedo]

Thank you for your interesting work, but I have questions about the results of TransMIL's comparative experiment in TCGA in the article. The data set is the data set of the DSMILused in your article, and the training set and test set are the same as in the article. The official code of TransMIL is here, the parameters used are also official parameters, and the results I ran were 93.3% accurate. How did 85.24% of the article come from

In addition, I also noticed the article SSLP: Spatial Guided Self-supervised Learning on Pathological Images of your team MICCAI2021 a few years ago, which said that the code would be made public, but it has not been made public at present

Can you explain that? Thank you！

Hi， We use the patches released by dsmil and extract the feature following dsmil. So we guess the different results mainly come from the different splits. We do not split the dataset like the TransMIL format, however we keep the same splits for all the feature extractor and aggregators, which is a fair comparison.

For SSLP, we have extended it into a journal paper SGCL (accepted by MedIA), and the code was also released.

[HHHedo]

BTW, we reproduce all the compared methods instead of directly citing their results, so there is no worry about a fair comparison like that issue. And you may also try interventional training on your split. Is there anything we should further explain?

[lihua8848]

Thank you for your interesting work, but I have questions about the results of TransMIL's comparative experiment in TCGA in the article. The data set is the data set of the DSMILused in your article, and the training set and test set are the same as in the article. The official code of TransMIL is here, the parameters used are also official parameters, and the results I ran were 93.3% accurate. How did 85.24% of the article come from In addition, I also noticed the article SSLP: Spatial Guided Self-supervised Learning on Pathological Images of your team MICCAI2021 a few years ago, which said that the code would be made public, but it has not been made public at present Can you explain that? Thank you！

Hi， We use the patches released by dsmil and extract the feature following dsmil. So we guess the different results mainly come from the different splits. We do not split the dataset like the TransMIL format, however we keep the same splits for all the feature extractor and aggregators, which is a fair comparison.

For SSLP, we have extended it into a journal paper SGCL (accepted by MedIA), and the code was also released.

Thank you for your patient answer, it is really helpful. But i still have some questions. In our experiment, we used the following pre-processed features, and I found this issue, he used self supervised training (simclr) to extract features, so I think this is the reason for such a big difference in our results.

---

However, according to your answer, do you use dsmil to cut patches from the original image and then use resnet18 to extract the feature?

1. Download the TCGA-NSCLC dataset
2. Cut the patches using dsmil $ python deepzoom_tiler.py -m 0 -b 20
3. Use resnet18 to extract the feature $ python compute_feats.py --dataset=TCGA-lung-single

I would like to know the specific parameters about cutting patches and using resnet18 to extract features? Thank you again for your work and answers.

[HHHedo]

issue

As mentioned before, we use the tcga-nsclc patches released by dsmil. Please refer to their great work.