thoroughimages / camel Goto Github PK

I want to learn more about this project.

When and how get total code?

Thanks very much!

Hi!
Thank you for your previous support! I'm currently having a question about generating dataset for retrain phase, hoping to get the clarification from you.

When combining the predictions of cMIL models to have a training dataset for retrain, what should we do if the models produce contradicting prediction results for a particular patch/ instance (e.g. Max-Max predicts a patch as NC while Max-Min predicts as CA)? Let's say if these results aren't considered as confusing patches. Do we include that patch in both NC & CA class in the training set?

Thank you so much! Look forward to hearing from you

Does the instance level classification performance(Table 4) referred to the classification performance on tiles or images? if tiles, perhaps extra pixel-level annotation on adenoma dataset are required to calculate it.

Hi,

Do you have a plan to open-source this project?

I think your work is contributive.

Best

Hi
Thanks for providing this great paper and dataset.

1. I found that the training of the MIL model need spend a lot of time because before the instance selection in every step you must inference the whole tiles in your training set how do you deal with it ?
2. At the beginning of the training of MIL, the weight of the model are random so the selection of the tile is also random. Will you pretrain the model in the noisy dataset(Labeling each tile with the Slide's label) then train the MIL model to make the selection of the tile at beginning of MIL training procedure meaningful ?

Hi!
Thank you so much for the paper and dataset. I'm currently implementing CAMEL following the procedure in the paper and I have some unclear thoughts, hoping you could help me clarify those.

1. About Max-Max & Max-Min, since they focus on the cancerous prediction of patches within an image, I select the appropriate patch, for each image, for the backprop by only considering the max/ min value of cancerous outputs from the model (e.g. I have 2 patches with NC-CA predictions from the model are (0.4, 0.6) and (0.2, 0.7) --> I choose the second patch for backprop). I wonder if this is a reasonable implementation or is there anything misunderstood?
2. In the paper, you mentioned at the end of 3.1.1 section that "Noted that we discard those potentially confusing samples whose predicted labels are different from their corresponding image-level labels." I infer that they are those with NC labels but predicted as CA. Could you please explain a bit more on what "confusing samples" mean?
3. I wonder if instance-level labels are the same as patch-level labels (i.e. we split an image into several smaller patches (e.g. 320x320))?

Looking forward to your reply. Thank you!

百度网盘给的数据有点问题，在解压的时候有部分数据出现问题，这是什么原因？

Hi.

Thanks for the great work. Will you host the dataset on a different platform in future? Currently, without a Chinese phone number or ID , I am unable to download your dataset. Any recommendations on how to download the data in this case?

Thank you

Hello, Thanks for your responses several years ago. I have implemented your paper these days, but I also have some questions about the implementation details.

1. In cMIL training part, I know you use 4 single GPUs(1 image for 1 GPU) to average their losses. Since I only have one GPU with 12GB memory, I just shrink the original image to the half and use the batch size of 4 to simulate your training step. I wanna ask how to arrange images into the batch, I use 2 positive images and 2 negative images for one batch, is this same as yours?
2. In Label Enrichment Part, Do you split images into training, validation and testing dataset and use validation dataset to find the convergence condition? Or just use training curve to get the lowest loss? I mean do you consider overfitting in cMIL training part? or overfitting is not necessary because we only need to generate pesudo labels of all images.
3. How to evaluate the performance of cMIL training? I use accuracy, precision, recall, confusion matrix to evaluate it. If the prediction value of only one patch of the whole image is larger than 0.5(due to sigmoid activation function in the last layer), I recognize it as positive. In other words, prediction values of all patches are smaller than 0.5 and then I regard this image as negative. However, in this metric, my result is so bad, which is around 50% accuracy.

Since I plan to use your work as a benchmark in my thesis paper, your comments are important for me. Looking forward to your reply.

Best,
Lingfeng

Thanks for your excellent work, recently I am re-implementing your work but meet some problems.

Have you used batch normalization in cMIL and retrain steps, since the batch arrangement is different in cMIL and in retrain step.

Looking forward to your reply.

Thank you.