Hi, I was trying to reproduce your results, but I see that the code for TCGA-NSCLC and TCGA-RCC datasets is missing. Could you add missing experiment setups?

Hello,

I randomly select 90% and 10% of official train data as train-set and val-set. The test-set is official provided. But the gap between training and inference is big. Have you encountered this problem?

As the dataset used in this paper is different from the compared baseline methods, how their performance results are obtained? Did you train and eval the baseline models respectively? Thanks.

Some WSIs have only 2048 patches, some have over 40000 patches. In your paper, you used a Squaring method to pad the embeddings to a fixed length. What is the fixed length? If the largest WSI has 50000 patches, did you pad all the embeddings to (1, 50000, 1024) ?

According to your paper, the total number of TCGA-NSCLC slides is 993 and you used 4-fold cross-validation, so each slide is classified once in hold-out sets after the cross-validation. By checking the numbers in table 1 from your paper: None of these accuracy values can be obtained by dividing an integer using 993. (i.e., N/993=acc does not resolve to an integer for the number of correctly classified slides N). Could you give an explanation on this?

Hi, thank you for your sharing!

I met some problems when I tried to use multi gpus to train, which was seemly caused by pytorch-lightning version:

pytorch_lightning.utilities.exceptions.MisconfigurationException: You have asked for amp_level='O2' but it's only supported with amp_backend='apex'

init() got an unexpected keyword argument 'dirpath'

Can you tell me your pytorch-lightning and some other packages' version? Thank you very much!

Hello,

Really fantastic work, thank you for the excellent repo.

I want to apply this to my own dataset and would love to produce attention heatmaps like you did in the paper. I understand that the Nystromformer has a return_attn argument however I am confused by the dimensions that it returns. I played around with a toy dataset that had 1000 instances and it returned a 1 x nheads x 1280 x 1280 tensor. Confused how to take that 1280x1280 array and tease out the cls_token attention values.

Any advice?

Thanks so much!

The directory of the camelyon16 dataset I downloaded is as follows. Where should I download the .pt files?

CAMELYON16
  -training
     -normal
        -normal_xxx.tif
     -turmor
        -turmor_xxx.tif
     -lesion_annotation
        -turmor_xxx.xml
  -testing
     -images
        -test_xxx.tif
     -evaluation
        -evaluation_masks.zip
        -evaluation_matlab.zip
        -evaluation_python.zip

Hello,

Does it seem that you haven't mentioned how to compute the attention of tokens in the paper? And I also can't find the code to compute the heatmap and visualization.

Can you answer my questions? Looking forward to your replay.

Hi all,

Thanks for your work. All looks good except I have been trying for some time to reproduce your results. I just discovered the thread https://github.com/szc19990412/TransMIL/issues/4 where others also have failed to reproduce the results. More specifically, I have been trying to reproduce the CAMELYON16 TransMIL ablations and the ABMIL benchmark using your repository and my own repository (written using PyTorch Lightning).

Some things I would like to know for reproducibility of the experiments:

• The given CAMELYON16 fold is only one fold, based on a 5:1 train-val split. In your paper you report using 10:1 train-val splits. I would therefore like to know what splits you used and whether they were stratified or not. For reproducibility I would like to request all splits to be uploaded to the repository so people are working with the same data.

• The CAMELYON16 dataset contains slides that were sourced from two centers that used different magnifications for each scanner (https://jamanetwork.com/journals/jama/fullarticle/2665774). If 20x magnification was used for the whole dataset, could you please confirm whether the slides from these two centers were tiled at two different physical resolutions (microns per pixel)? “The whole-slide images were acquired at 2 different centers using 2 different scanners. RUMC images were produced with a digital slide scanner (Pannoramic 250 Flash II; 3DHISTECH) with a 20x objective lens (specimen-level pixel size, 0.243 μm × 0.243 μm). UMCU images were produced using a digital slide scanner (NanoZoomer-XR Digital slide scanner C12000-01; Hamamatsu Photonics) with a 40x objective lens (specimen-level pixel size, 0.226 μm × 0.226 μm).”

• The model selection/checkpointing details. I see from the code https://github.com/szc19990412/TransMIL/blob/3f6bbe868ac39e7d861a111398b848ba3b943ca8/utils/utils.py#L43-L49, that you used EarlyStopping with patience=10 on val_loss, correct?

• What optimizer and hyperparameters were used for ABMIL? I assume what the authors (https://arxiv.org/pdf/1802.04712.pdf) report in Table 17 for the histopathology datasets?

I will later report my scores for TransMIL & ABMIL using the repository and given data fold. Thanks in advance.

Hi, I was wondering how large slides were handled during inference and training. Was there any limit on the bag-size to prevent OOMs? If so can you clarify how the predictions from multiple bags were aggregated.

Thanks in advance.

The codes worked really well on the new dataset.

Is it possible to provide the codes for visualization? I know you used some codes from CLAM, but I still unclear of how you generate the mappings in detail.

Thank you so much!

i would like to ask how much the training step takes per Epoch, i used your built model and i modified the PPGE model by adding FFT to reduce the dimension Convolution operation, only issue i noticed was that the Trainer took a lot of time to finish single Epoch, that's is related to size the shape of the image (2154,1024) or i missed something