ahmedbesbes / mrnet Goto Github PK

Building an ACL tear detector to spot knee injuries from MRIs with PyTorch (MRNet)

License: MIT License

Python 100.00%

pytorch medical-imaging mri-images mrnet convolutional-neural-networks cnn data-augmentation transfer-learning computer-vision pytorch-tutorial

mrnet's Introduction

Deep learning in medical imaging: How to automate the detection of knee injuries in MRI exams ?

This repository contains an implementation of a convolutional neural network that classifies specific knee injuries from MRI exams.

It also contains the matieral of a series of posts I wrote on my blog.

Dataset: MRNet

The data comes from Stanford ML Group research lab. It consits of 1,370 knee MRI exams performed at Stanford University Medical Center to study the presence of Anterior Cruciate Ligament (ACL) tears.

For more information about the ACL tear problem and the MRNet data please refer to my blog post where you can investigate the data and build the following data visualization in jupyter notebook:

To learn more about the data and how to realize this visualization widget, read my first post.

Code structure:

This charts summarizes the architecture of the project:

For more details about the code, please refer to my second blog post .

How to use the code:

If you want to retrain the network on your own you have to ask for the data from Stanford via this link.

Once you download the data, create a data folder and place it at the root of the project. You should have two folders inside: train and valid as well as a bunch of csv files.

To run the script you can execute it with the following arguments:

parser = argparse.ArgumentParser()
parser.add_argument('-t', '--task', type=str, required=True,
                    choices=['abnormal', 'acl', 'meniscus'])
parser.add_argument('-p', '--plane', type=str, required=True,
                    choices=['sagittal', 'coronal', 'axial'])
parser.add_argument('--augment', type=int, choices=[0, 1], default=1)
parser.add_argument('--lr_scheduler', type=int, choices=[0, 1], default=1)
parser.add_argument('--epochs', type=int, default=50)
parser.add_argument('--lr', type=float, default=1e-5)
parser.add_argument('--flush_history', type=int, choices=[0, 1], default=0)
parser.add_argument('--save_model', type=int, choices=[0, 1], default=1)
parser.add_argument('--patience', type=int, choices=[0, 1], default=5)

example to train a model to detect acl tears on the sagittal plane for a 20 epochs:

python -t acl -p sagittal --epochs=20

Note: Before running the script, add the following (empty) folders at the root of the project:

models
logs

Results:

I trained an ACL tear classifier on a sagittal plane and got the following AUC scores:

on train: 0.8669
on validation: 0.8850

Logs on Tensorboard:

Contributions - PR are welcome:

If you feel that some functionalities or improvements could be added to the project, don't hesitate to submit a pull request.

mrnet's People

Contributors

Stargazers

Watchers

mrnet's Issues

An error occurred while running.：AttributeError: Can't pickle local object 'load_data.<locals>.<lambda>'，How do you solve this problem，thank you very much

part 3: Interpret models' predictions

Hi, thank you for your wonderful work. I'm trying the third part: Interpret Models' Predictions. I find I can't visualize some of the predictions. As follows:

IndexError Traceback (most recent call last)
in
1 import shutil
----> 2 create_patiens_cam(acl[0],plane)

in create_patiens_cam(case, plane)
19 logit = mrnet(mri)
20 size_upsample = (256, 256)
---> 21 feature_conv = feature_blobs[0]
22 # print(features_blobs[0])
23

IndexError: list index out of range

and then I did this:
mrnet = torch.load(f'../models/{model_name}')
mrnet = mrnet.to(device)
mrnet.eval()

hook the feature extractor

features_blobs = []
def hook_feature(module, input, output):
features_blobs.append(output.data.cpu().numpy())
mrnet._modules.get('pooling_layer').register_forward_hook(hook_feature);

...

global features_blobs
mri = mri.to(device)
logit = mrnet(mri)
size_upsample = (256, 256)
feature_conv = features_blobs[0]
h_x = F.softmax(logit, dim=1).data.squeeze(0)
probs, idx = h_x.sort(0, True)
probs = probs.cpu().numpy()
idx = idx.cpu().numpy()
slice_cams = returnCAM(features_blobs[-1], weight_softmax, idx[:1])

...

import shutil
create_patiens_cam(acl[0],plane)

Could you tell me what went wrong? thank you

getting an error while running the code.

can you please help me solve the error.

Seems that "Softmax" should be used instead of "Sigmoid"?

Hello, thanks for your implementation. However, I found that the "probs" for binary classification doesn't sum up to be 1.0.


        prediction = model.forward(image.float())
        loss = torch.nn.BCEWithLogitsLoss(weight=weight)(prediction, label)
        loss.backward()
        optimizer.step()

        loss_value = loss.item()
        losses.append(loss_value)

        probas = torch.sigmoid(prediction)

        y_trues.append(int(label[0][1]))
        y_preds.append(probas[0][1].item())

The issue mentioned is located in "https://github.com/ahmedbesbes/mrnet/blob/master/train.py"

I can not reproduce your train AUC, & valid AUC scores

Dear Besbes,
Firstly, I thank for your sharing code. It is very nice. I test with sagittal dataset with anterior cruciate ligament tear. However, I could not reproduce your results like train & valid AUC scores. So could you explain to clarify the issue? Or maybe I acquire a mistake somewhere?
Best regard. Linh