search

nyukat / breast_cancer_classifier

Deep Neural Networks Improve Radiologists' Performance in Breast Cancer Screening
https://ieeexplore.ieee.org/document/8861376
GNU Affero General Public License v3.0
840 stars 269 forks source link

Training procedure? #33

Closed waleedrazakhan92 closed 4 years ago

waleedrazakhan92 commented 4 years ago

Hi! I would like to train your model on a dummy dataset. Can you guide me or better yet provide some helper code.? Specifically for the view wise model for a single image. The model outputs a csv file with the resulting probabilities for malignant/benign, so i'm assuming that the training dataset would be images with binary labels for malignant and benign classes. Am I assuming right?

kjgeras commented 4 years ago

Hi,

We are currently not working on releasing our training code.

Yes, if you use the model that is learning with one image at a time (this model https://github.com/nyukat/breast_cancer_classifier/blob/master/using_tensorflow.md) then the training data set would consist of (image, malignant_label, benign_label) triples.

waleedrazakhan92 commented 4 years ago

Hi,

Thank you for your response. I also noticed in the sample .ipynb you provided, when you load the model you specifically set the view parameter as 'L-CC', which implies that you're loading 'L-CC' model. And also during inference you give an 'L-CC' view image, which obviously makes sense. So if i want to run inference on 'R-CC' or 'L-MLO' view image, do i have to load that view-specific model for each specific image?

Another question is that when i see the summary of that model (L-CC) model, it mentions output_features=4, and also when the inference is run on the model it outputs 4 values which you manipulate to get two output probabilities for malignant and benign. I didn't understand the very line ('np.exp(y_hat.cpu().detach().numpy())[:, :2, 1]') which converts those 4 tensor values of 'yhat' into the malignant/benign probabilities.

Also, lets say i want to train this model on my own dataset, do i have to train 4 models for each view separately? and then somehow combine the results of those 4 models? and also would each model would train on only view specific images (like the L-CC model would only run on L-CC images and R-CC would train on R-CC images and so on)?

kjgeras commented 4 years ago

Thank you for your response. I also noticed in the sample .ipynb you provided, when you load the model you specifically set the view parameter as 'L-CC', which implies that you're loading 'L-CC' model. And also during inference you give an 'L-CC' view image, which obviously makes sense. So if i want to run inference on 'R-CC' or 'L-MLO' view image, do i have to load that view-specific model for each specific image?

Yes, as far as I remember.

Another question is that when i see the summary of that model (L-CC) model, it mentions output_features=4, and also when the inference is run on the model it outputs 4 values which you manipulate to get two output probabilities for malignant and benign. I didn't understand the very line ('np.exp(y_hat.cpu().detach().numpy())[:, :2, 1]') which converts those 4 tensor values of 'yhat' into the malignant/benign probabilities.

I don't understand what exactly you are asking. Please follow what the model does in the example we provided.

Also, lets say i want to train this model on my own dataset, do i have to train 4 models for each view separately? and then somehow combine the results of those 4 models? and also would each model would train on only view specific images (like the L-CC model would only run on L-CC images and R-CC would train on R-CC images and so on)?

Yes, that is what you should do if you wanted to train that model. I think this model https://github.com/nyukat/GMIC is actually probably better for what you are trying to do.