invoker-ll / wsi-finetuning Goto Github PK

I noticed that you are using the BRCA dataset in the dataset, so I have a few questions to ask.

1. Has your BRCA dataset been subjected to common subtype binary classification (IDC and ILC).
2. If so, how did you map the labels of IDC and ILC from the BRCA dataset.

Hi, I encountered a problem while trying to run extract_topK_ROIs.py. After debugging, I think the issue lies in the self.wsi.read_region function. Upon checking the OpenSlide documentation, it appears that the coordinates should be specified as x and y. However, when I printed out the coordinates, I found only a single value (e.g., 108736). Could you help me overcome this issue? Thank you!

Hi, I am confused about the vib patch-selection process in the code. In the paper it mentioned that the patches are selected based on the bernouli probability, but I found that in the code it is actually selected based on the attention scores from CLAM. I tested the code and the performance is not good. Could you help me to solve this? Thanks.

logits, Y_prob, Y_hat, A, results_dict = model_wsi(features, testing=True)
topk_indexes = torch.topk(A.squeeze(), min(args.topk_num,A.shape[1]), sorted=True)[1].cpu()

Hello, thank you for your code. However, I met some problem in Stage-2 (wsi-finetuning with topK).

In extract_topK_ROIs.py line 88,

dataset = Whole_Slide_Bag_FP_SAVE(file_path=patch_file_path,wsi=wsi,select_idx=topk_indexes)

However, the class Whole_Slide_Bag_FP_SAVE doesn't have the keyword argument select_idx. Could you please provide the complete code?
Thanks a lot!

Hello, I'm confused about the mask operation in reducing patches.
In the implementation, the mask is obtained by addition of z_mask (sampling from Bernoulli distribution) and instance_prob (predict by linear layer), I think it should be mask = z_mask. could you please tell me the reason you compute the mask in this way? Thanks a lot.

Hi, I tried following your code to create patches by using bash create_patches.sh. However, there are two missing slides (normal_027 and normal_045), and as a result, I could only get 397 instead of 399 on Camelyon16 dataset. Do you perhaps know why?

Thank you!

Hi, thanks for your interesting work. I have a question about your loss function (Eq.(8)) in your paper.
In Eq.(8), you have stated that we should use KL loss in the latter part. However, I didn't see any KL loss code in your code (All I have found is MSE loss ). Could you please tell me where did you use your KL loss in your code? Thanks.

Hi, I wonder if you have the complete code in which you combine FT with other methods such as Trans-MIL, DFTD-MIL, Mean-pooling, Max-pooling, and KNN, as shown in the paper. As I believe the current code is only applicable to CLAM.

Thank you!

    # temp_split=i
    temp_split = 0 # to avoid validation data leakage

https://github.com/invoker-LL/WSI-finetuning/blob/e1c6972a3a1fa47b2e3497fb042cbc8a6372431d/train_end2end.py#L39C8-L40C58