Hello,
I achieved very good and very fast inference with your CT model. Are the MRI weights available somewhere?
thx e.-

After I couldn't get the notebook to work on my own caffe installation (see #29) I tried running your prepared docker. However this line

# Load network
net1 = caffe.Net(STEP1_DEPLOY_PROTOTXT, STEP1_MODEL_WEIGHTS, caffe.TEST)

throws the following error

...
F0509 19:58:49.805807    24 cudnn_relu_layer.cpp:15] Check failed: status == CUDNN_STATUS_SUCCESS (1 vs. 0)  CUDNN_STATUS_NOT_INITIALIZED

My host system is ubuntu 18.04, NVIDIA Driver Version 390.48, CUDA Version 9.1.85 (if that is any help)
caffe is set to CPU mode in your example jupyter notebook

Any help would be appreciated.

step2_weights.caffemodel I cannot find

why use class weight 1, 13,120 ? Can we use other number for weights ?...On what basis you are using this weights..

class_loss_weights: 1
class_loss_weights: 13
class_loss_weights: 120

Hi, I found your model link is disable, can you update your link or give me a web link, thank you

Hello,

I'm following this notebook: https://github.com/mohamed-ezz/saratan/tree/master/notebooks for Step 1 caffemodel, I just need clarification on my understanding once I trained for dataset 3D-Ircadb using this notebook so basically my model is ready to segment liver. How about for lesion, do I need to save the result after segmenting in Step1 and use the same notebook for creating the Step 2 model.
Please help me to understand, I don't see how's this is taken care in the code.

The illustration in the paper uses the 86th slice of the patient No. 9 in the data set,and I successfully ran the cascaded_unet_inference.ipynb program in Docker,but the result is inconsistent with the picture in the paper, where is the implementation of my right? Is not using 3DCRF? In the Docker how should we use denseinference?Looking forward to your answer

hi

I could only find the deploy proto
can you please provide the complete set of proto? train, train val?

thanks

Hi, I have tested your released model, my code is refer to your notebook. my code is

`caffe.set_mode_gpu()
caffe.set_device(2)
net_liver = caffe.Net('/home/zhou/zou/Cascaded-FCN/models/cascadedfcn/step1/step1_deploy.prototxt', '/home/zhou/zou/Cascaded-FCN/models/cas cadedfcn/step1/step1_weights.caffemodel', caffe.TEST)

img=read_dicom_series("../train_image/3Dircadb1.17/PATIENT_DICOM/")
lbl=read_liver_lesion_masks("../train_image/3Dircadb1.17/MASKS_DICOM/")
S = 90
img_p = step1_preprocess_img_slice(img[...,S])
lbl_p = preprocess_lbl_slice(lbl[...,S])
net_liver.blobs['data'].data[0,0,...] = img_p
pred = net_liver.forward()['prob'][0,1] > 0.5
plt.figure(figsize=(3*5,10))
plt.subplot(1, 3, _1)
plt.title('CT')
plt.imshow(img_p[92:-92,92:-92], 'gray')
plt.subplot(1, 3, 2)
plt.title('GT')
plt.imshow(lbl_p, 'gray')
plt.subplot(1, 3, 3)
plt.title('pred')
plt.imshow(pred, 'gray')`

but the result is very bad like this

Hi, I have run the notebook and seemed to obtain an incorrect prediction after step1.

For comparison, I have used the same example from https://github.com/cjayanth95/liverseg/blob/master/custom_outs_updated.ipynb.

This is my output: https://github.com/XiaonanZang/liverseg/blob/master/cascaded_unet_inference-Copy1.ipynb.

I have followed your instruction and built the caffe from the provided git.

Could you help me on this issue?

Thank you very much.

Xiaonan Zang

Hello,

I'm trying to reproduce your results and I'm having trouble with the CRF step. Following the ipython notebook, I'm using image 17 from 3DIRCAdb1. I've made some modifications in the notebook to run the network for every slice. Without CRF, I got a dice score of 0.8573.

I cloned and installed https://github.com/mbickel/DenseInferenceWrapper/. However, I'm not sure how I can use it. I'm trying the following:

pro = CRFProcessor.CRF3DProcessor()
result = pro.set_data_and_run(img_preprocessed, extended_labels)

Where img_preprocessed is the preprocessed image without the padding reflections (dimensions (W,H,D); and extended_labels is the output of the predictions (direct probabilities, not thresholded bool) with dimensions (W,H,D,L) (where L = 1). I'm getting a (W,H,D) matrix with zeroes.

Can you help me with this?

Hi,
I have abdominal CT scan data sets for 15 patients. Among them 5 data sets are with HCC legions. I also have the ground truths of liver for all the data sets and also ground truth of legions for the 5 data sets with HCC.
Can you help me train this data to create my own caffe model?

Thanks,
Jayanth.

You have reported a dice score of 82.3% for the lesions in the first paper. It's not clear as to what this metric belongs to. Is it the dice score for lesion at training or testing?

In the follow-up paper, it's been reported as follows: "In general, we observe significant additional improvements for Dice overlaps of liver segmentations, from mean Dice 93.1% to 94.3% after applying the 3D CRF. For lesions we could achieve a Dice score of 56% at a standard deviation of 26% with a 2 fold cross-validation."

Can you clarify me regarding the correct results? Why is there a discrepancy?

Hi, I have read your paper many times and I have some questions. Hope you can help me, thank you!
in Table 1, are all the results for liver segmentation? If the results are for liver, in my option the first branch of Cascaded UNET is for liver segmentation and it is just a UNET, why the result is very different from the UNET?(DICE 93.1% vs 72.9%).

This is my code:
STEP1_DEPLOY_PROTOTXT = "/home/mi/smj/myproject/Cascaded-FCN-master/models/cascadedfcn/step1/step1_deploy.prototxt"
STEP1_MODEL_WEIGHTS = "/home/mi/smj/myproject/Cascaded-FCN-master/models/cascadedfcn/step1/step1_weights.caffemodel"
STEP2_DEPLOY_PROTOTXT = "/home/mi/smj/myproject/Cascaded-FCN-master/models/cascadedfcn/step2/step2_deploy.prototxt"
STEP2_MODEL_WEIGHTS = "/home/mi/smj/myproject/Cascaded-FCN-master/models/cascadedfcn/step2/step2_weights.caffemodel"
img=read_dicom_series("/home/mi/smj/myproject/Cascaded-FCN-master/test_image/3Dircadb1.17/PATIENT_DICOM/")
lbl=read_liver_lesion_masks("/home/mi/smj/myproject/Cascaded-FCN-master/test_image/3Dircadb1.17/MASKS_DICOM/")

This is my result:

Why is result so bad?

Do you have the train.prototxt to finetune the attached models?

OS: ubuntu 16.04
GPU: Nvidia GeForce GTX 105
The kernel dies while loading network when working from docker
and when I set up my own environment I am able to run it through some layers only from net.forward() method but unable to get to the output, the kernel dies on forward pass. kindly suggest how to solve the issue as I am testing it only for one image right now as given in the notebook. Thank you

Hi, I studied u-net prototxt and your unet-overfit-python.prototxt . I found that you didn't use batch normalization layer. Can you explain why?

In step1, the dataset contains 15 patients to train liver. And for step2, do you still use these 15 patients' CT scans? You crop the liver and resized and padded to fit the u-net. Then I thought there are two problems:
1, for each slice, liver size is different, if you crop just the tiny liver, comparing with those "big" liver, there will be noises. In Medical image processing, they often crop the liver by the volume size, not the liver size of a single slice.
2, for 15 patients, 2063 slices in total, 1572 of them have livers, and only 579 of them have tumors. You just crop all livers to train lesion in step2? Then will there be imbalance problem?

I'm reading codes in https://github.com/IBBM/Cascaded-FCN/blob/master/notebooks/cascaded_unet_inference.ipynb and I feel confusing in step1_preprocess_img_slice, this bold comment:
6- Pad img slices with 92 pixels on all sides (so total shape is 572x572)
I wonder know why this preprocess needs to pad image with 92 pixels. I didn't find any reasons in paper. Please tell me. Thanks.

Hi there,

I was wondering what the format of the data was for you when you did the training of this model? I understand the preprocessing steps for the example, but when the training happened was it loaded in through .npy files? I was trying to figure it out by looking through https://github.com/mohamed-ezz/saratan/blob/master/notebooks/realtime_train.ipynb but was having issues. Thanks for your help.

The dropbox link of the two weights.caffemodel were expired, the dropbpx told me the file was not exist or has been moved.

RuntimeError: Could not create write struct with imshow

which library to use for imshow - opencv or matplotlib?

Hi,

I would like to train my dataset , which consist of ultrasound 2D images, in gray-scale, for making automatic bone segmentation . I have no GPU, so I would like to finetune your model in order to do the training faster . My images are of different sizes( the height is fixed at 512, but the width varies between 300 and 512. Is there any special consideration in this case ? Your cand find attached an example of an image and the ground truth

Thank you very much,

Cordially,

Mateo Villa

Hi,
I am having trouble understanding pre-processing step. I have dicom images of liver of size 512x512. How do I proceed further?Do I have to resize it to 388x388 and add symmetrical padding of size 92?

Any help would be appreciated.

Thanks.

Firstly, let me commend you on your effort for this paper and the repo that accompanies it. As a programmer myself, I know how difficult it is to write code especially for machine learning models.

I successfully, reproduced the results through a GPU hardware and the code code ran up until:

# Visualize results imshow(img_p, lbl_p, pred>0.5, title=['Slice','Ground truth', 'Prediction'])

Besides, I also had to downgrade SCIPY to 1:00 because of the depreciated imgresize issue.

/usr/local/lib/python3.6/dist-packages/ipykernel_launcher.py:108: DeprecationWarning: imresizeis deprecated!imresizeis deprecated in SciPy 1.0.0, and will be removed in 1.2.0. Use ``skimage.transform.resize`` instead. /usr/local/lib/python3.6/dist-packages/ipykernel_launcher.py:104: DeprecationWarning:imresizeis deprecated!imresize is deprecated in SciPy 1.0.0, and will be removed in 1.2.0. Use ``skimage.transform.resize`` instead.

perhaps you could look into these issues.

Additionally, running:
`# Prepare liver patch for step2

net1 output is used to determine the predicted liver bounding box

img_p2, bbox = step2_preprocess_img_slice(img_p, pred)
imshow(img_p2)`

gave the following error:
`---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
in ()
----> 1 img_p2, bbox = step2_preprocess_img_slice(img_p, pred)
2 imshow(img_p2)

in step2_preprocess_img_slice(img_p, step1_pred)
79 y2 = min(img.shape[0], y2+y_pad)
80
---> 81 img = img[y1:y2+1, x1:x2+1]
82 pred = pred[y1:y2+1, x1:x2+1]
83

TypeError: slice indices must be integers or None or have an index method`

these affected the rest of the executions such as:

1. `# Visualize result

extract liver portion as predicted by net1

x1,x2,y1,y2 = bbox
lbl_p_liver = lbl_p[y1:y2,x1:x2]

Set labels to 0 and 1

lbl_p_liver[lbl_p_liver==1]=0
lbl_p_liver[lbl_p_liver==2]=1
imshow(img_p2[92:-92,92:-92],lbl_p_liver, pred2>0.5)`

2. # Load step2 network net2 = caffe.Net(STEP2_DEPLOY_PROTOTXT, STEP2_MODEL_WEIGHTS, caffe.TEST)

and finally

3. net2.blobs['data'].data[0,0,...] = img_p2 pred2 = net2.forward()['prob'][0,1] print (pred2.shape)

Please could you kindly proffer solutions to these issues?

Can you provide the pre-trained tensorflow model(including graph and ckpt) for us? Training from scratch really takes a lot of time and need gpus with big memory. Thank you very much!

There are some irrelevant images (without liver region) in the CT volumes. Do I need to remove these images or send all of them to the network as a training set?
For metrics, I didn't get any prediction for small image with very small liver region. So some of the metrics cannot be calculated (Infinite). Do you get a predicted result for every single image, and compared with the ground-truth of foreground?

Hello, could you provide your trainning code for us?Thanks very much.

Hi, I execute the cascaded_unet_inference.ipynb that doesn't report any questions, but the result is worse.
The result of step1 is shown as follow:

Any one know the reason?
Thank you very much

On which dataset training is done and also how testing is performed (name of both datasets would be fine)

Also could you pls highlight the evaluating function

Hello!

If I understand right your input images size is increased from 388x388 to 572x572, the same dimensions as in the U-Net publication by Olaf Ronneberger, by extending all sides with mirrored copies of the image.

But since the background of the CT scans has intensities near zero anyway, wouldn't it be possible to use zero padding instead and remain at the original size? The network should then be able to use less memory and to return a label map of the same size as the input, so I wonder if you found a reason not to use the zero padding that I have been missing so far.

Model Address invalidation?
Could you update the address again?