The link to the ISIC2018 dataset is not accessible, can you provide an up-to-date link？

I already saw your code and your answer about other person's issue.
"
you can use the following code to include bi-directional ConvLSTM:
LSTM_f = layers.ConvLSTM2D(filters = 128, kernel_size=(3, 3), padding='same', return_sequences = False, go_backwards = False, kernel_initializer = 'he_normal')(X)
LSTM_b = layers.ConvLSTM2D(filters = 128, kernel_size=(3, 3), padding='same', return_sequences = False, go_backwards = True, kernel_initializer = 'he_normal')(X)
Bi_rep = layers.Add()([LSTM_f, LSTM_b])
"

There is no Bi-direction convLSTM2d. so If i want to use bi-directional convlstm2d, Do I just typo that your code LSTM_f and LSTM_b ??

Maybe it's the problem of experimental setup. My experimental results are worse than those provided in the paper. Is there any extra tricks to follow?

I tried to run "python prepare_datasets_DRIVE.py" and got the error.

original image: 21_training.tif
tempfile.tif: Cannot read TIFF header.
Traceback (most recent call last):
File "prepare_datasets_DRIVE.py", line 82, in
imgs_train, groundTruth_train, border_masks_train = get_datasets(original_imgs_train,groundTruth_imgs_train,borderMasks_imgs_train,"train")
File "prepare_datasets_DRIVE.py", line 46, in get_datasets
imgs[i] = np.asarray(img)
TypeError: float() argument must be a string or a number, not 'TiffImageFile'

Dear Reza,

We would like to use some of your project's features in an application we plan to release under the permissive open source Apache v2.0 license, but I found no copyright or license notice in the project repository. Can you clarify this information and make it appear in your source code? Without it, it is impossible to legally use your source code, because copyright law strictly applies and prohibits any reuse.

If you don't have any idea about this, you'll find detailed information in the Github documentation and on the Open Source Guides:

• Licensing a repository
• The Legal Side of Open Source

Thank you very much for your work and thank you in advance for your clarifications.

Sébastien

I noticed you are flattening the whole batch like this.

This way, a pair of masks where the Dice Coefficient is close to 0 like below won't be penalized heavily, as the metric is being averaged over pixels, not over images.

i use the DRIVE dataset ，but why my training result value is the same。
#################################################
Epoch 1/50
160000/160000 [==============================] - 10087s 63ms/step - loss: 0.3565 - acc: 0.8929 - val_loss: 0.3562 - val_acc: 0.8855
Epoch 2/50
160000/160000 [==============================] - 11281s 71ms/step - loss: 0.3405 - acc: 0.8929 - val_loss: 0.3562 - val_acc: 0.8855
Epoch 3/50
160000/160000 [==============================] - 12124s 76ms/step - loss: 0.3405 - acc: 0.8929 - val_loss: 0.3561 - val_acc: 0.8855
Epoch 4/50
160000/160000 [==============================] - 15388s 96ms/step - loss: 0.3405 - acc: 0.8929 - val_loss: 0.3562 - val_acc: 0.8855
Epoch 5/50
160000/160000 [==============================] - 19182s 120ms/step - loss: 0.3405 - acc: 0.8929 - val_loss: 0.3563 - val_acc: 0.8855

######################

Could you provide "processed_data" file please? Thank you

Good day to you. Is it possible to pass more than single channel (grayscale) arrays to model. I would like to train on my own dataset with MRI images: T1-wrighted, T2-weighted, T2 TIRM, DWI, SWI. All of images have same shape (512, 512, 1) and want to encode them as (512, 512, 5) input.
Should I modify model exept input_size=(512, 512, 5) param of BCDU_net_D3 function?

P.S. Sorry for my english :(.

Hello, I try to use Retina Blood Vessel Segmentation. I would like to ask you what should I do when I run train_retina.py and have this error. I hope you can see my question.

while training for the skin images data set on Google Colab I came across errors in a data preparation section. where this line:
img = np.double(sc.imresize(img, [height, width, channels], interp='bilinear', mode = 'RGB'))

and this line:
img2 = np.double(sc.imresize(img2, [height, width], interp='bilinear'))

showed errors, as scipy.misc.imresize is deprecated in newer versions of scipy. What I did is I used skimage.transform.resize instead of scipy.misc.imresize, that's what is recommended, as:

img = np.double(resize(img, (height, width, channels)))

Here I have imported resize from skimage.transform
and
img2 = np.double(resize(img2, (height, width)))

It worked. But the problem is the data reading process has become very slow as compared to the scipy.misc.imresize. I had tested it before on old version of scipy.misc.imresize. Any help in this regard would be very much appreciated. Thanks!

Hello, sir!
I trained the model for the skin images, everything went smooth and well. In the evaluate.py this error occured:

I am training the model on Google Colab. Kindly look into it. I am looking forward for your response. Thanks a lot!
Asif Ahmad

Excuse me .Thanks once again for awesome work. when I learn the code ,I find it doesn't have model.compile ,but model.fit is work。can you help me to understand the loss and the optimizers?

i first run the skin lesion project and succeed
then i try to change my own dataset, it is .nii format, and i train model have the following picture

but when i evaluate a nii(and it is in my validation), but it evaluate the same value at every pixel, i have no idea how to fix it

can you help me please, thx!

Hi,

I was trying to implement lung segmentation from the repo. As mentioned, I downloaded the dataset from Kaggle. I processed the data and the npy files are now being saved inside the processed_data folder under Lung Segmentation. In order to train the network, I was trying to use train.py file, but ran into an error that states:

File "train_lung.py", line 20, in
tr_data = np.expand_dims(tr_data, axis=3)
File "<array_function internals>", line 6, in expand_dims
File "/home/achoudhury/biomedseg/lib/python3.6/site-packages/numpy/lib/shape_base.py", line 597, in expand_dims
axis = normalize_axis_tuple(axis, out_ndim)
File "/home/achoudhury/biomedseg/lib/python3.6/site-packages/numpy/core/numeric.py", line 1358, in normalize_axis_tuple
axis = tuple([normalize_axis_index(ax, ndim, argname) for ax in axis])
File "/home/achoudhury/biomedseg/lib/python3.6/site-packages/numpy/core/numeric.py", line 1358, in
axis = tuple([normalize_axis_index(ax, ndim, argname) for ax in axis])
numpy.AxisError: axis 3 is out of bounds for array of dimension 2

Can you please help me with it?

Thank you so much!

1: Skin Lesion dataset download link is not active.
2: In prepare_dataset_DRIVE.py of retina blood vessel segmentation folder this script is not running properly.

print(patches_imgs_test.shape)
(214200, 64, 64, 1)

predicted images size :
(214200, 64, 64, 1)

if average_mode == True:
    pred_imgs = **recompone_overlap**(pred_patches, new_height, new_width, stride_height, stride_width)# predictions
    orig_imgs = my_PreProc(test_imgs_orig[0:pred_imgs.shape[0],:,:,:])    #originals
    gtruth_masks = masks_test  #ground truth masks

else:
    pred_imgs = recompone(pred_patches,13,12)       # predictions
    orig_imgs = recompone(patches_imgs_test,13,12)  # originals
    gtruth_masks = recompone(patches_masks_test,13,12)**

**Traceback (most recent call last):

  File "C:\Users\BIOIMA~1\AppData\Local\Temp/ipykernel_25740/3778051638.py", line 2, in <module>
    pred_imgs = **recompone_overlap**(pred_patches, new_height, new_width, stride_height, stride_width)# predictions

  File "C:\Users\Bioimaging\shrirang_IPW\conference\data\DRIVE\BCDU-Net-master\Retina Blood Vessel Segmentation\extract_patches.py", line 266, in recompone_overlap
    assert (preds.shape[1]==1 or preds.shape[1]==3)  #check the channel is 1 or 3

AssertionError**

I think here the error is occuring because of this >> recompone_overlap

Train_img      = Data_train[0:alpha,:,:]
Test_img       = Data_train[alpha:en_d,:,:]

Train_mask     = Mask_train[0:alpha,:,:]
Test_mask      = Mask_train[alpha:en_d,:,:]

Train_maska     = Maska_train[0:alpha,:,:]
Test_maska      = Maska_train[alpha:en_d,:,:]

FOV_tr     = FOV_train[0:alpha,:,:]
FOV_te      = FOV_train[alpha:en_d,:,:]

in module Prepare_data.py

Dear author, thank you very much for opening up your code.
I want to get the visualization results after lung segmentation. After I read the DCM file with simpleitk library, the results are shown in the figure below

Obviously, this result is wrong.
model = M.BCDU_net_D3(input_size = (512,512,1))
model.load_weights('./weight/weight_lungD3.hdf5')
vol = sitk.ReadImage('IM000090.dcm')
volarray = sitk.GetArrayFromImage(vol)
pre = model.predict(volarray)

This is the code I use. How do you visualize it, or what do you use as the input of the model?

Can you tell me the meaning of Data_train, Mask_train, Maska_train，FOV_train obtained through Prepare_data.py?

In file "evaluate_performance.py ", you flatten all tesing data to an array, and fed it into jaccard_similarity_score,f1_score and so on. However, the general practice is to evaluate each input data and then average them. The results of the two methods are quite different, although the accuracy is the same.

In this code, is there still only one direction adopted by ConvLSTM?

As you have mentioned ConvLSTM2D is used for Spatio-temporal information.

But I am confused, is this model really going to learn the temporal information? There are two reasons for it.

• Your datasets are images and not videos.
• And you are using a Convolutional layer in your encoder and decoder but Bi-Directional ConvLSTM in skip connection.

Can you share your findings? Does this really work?

Reference:
Temporal Information
Stackoverflow answer

i use CT dataset, i found the label with value of 1 is less than 0.003 in whole CT picture ,the i use patch method.
i have 250 CT pictures,and size is 256256 , the i divide into 6464 , 220 CT into 198006464 patch for train, and 30 CT into 27006464 for valadation
i tried your D3model and Unet model, and i also tried binary_crossentropy loss, dice loss and GDL, first i use the learning rate 0.0001, but loss first declined but quickly increased in the first epoch
the i decrease lr like 0.0005, 0.00001, but it didn't work.
and i found your unet model only maxpooling 3 times, i tried the 32 layers Unet with 4 times maxpooling, i found lr 0.0001 for 256256 didn't work , i tried 0.00001 works.
so i don't konw how to select the lr for pathes of 6464 CT, i think enitial lr 0.00001 is small ,can you help me ,thx!

do I have to train the model? I just run the "evaluate.py" in folder "Skin Lesion Segmentation" and it reports an error:

FileNotFoundError: [Errno 2] No such file or directory: 'data_test.npy'

so , is this file generated during the trainning process?

Hi

I would like to run your pre-trained model on some retina images from a different dataset.

Your evaluate model seems to be dependent on some hdf5 structure. Is there a simple method you could propose to make segmentation prediction possible from some .png images?

Michael

I found that in Skin Lesion Segmentation task, you used 'jaccard_similarity_score' function to evaluate ‘JS’ score in your code. After version 0.23, sklearn used 'jaccard_score' instead of it, because "the current Jaccard implementation is ridiculous for binary and multiclass problems, returning accuracy"(scikit-learn/scikit-learn#13151). So your JS score may be inaccurate, after my verification, the actual score should be about 0.76

Hello sir,

Thank you very much for this great repository. I used pretrained models for DRIVE dataset but i received a wrong mask. I have done the image grayscale before input and then I take this result. Αm i doing something wrong?

Thank you in advance

Hi, Thanks for the interesting work. I was wondering how to extend this idea for multi-class segmentation.

1)jaccard_similarity_score from package sklearn.metrics used in your codes should be old and the results maybe wrong. After Sklearn Version 0.32 , jaccard_similarity_score is replaced by jaccard_score, the testing results become lower. When binary segmentation, jaccard_similarity_score == Accuracy.
2)The unet model is hard to train and there is no validation accuracy improvment when training on my GPUS:
815/1815 [==============================] - 79s 44ms/step - loss: 0.6670 - acc: 0.7923 - val_loss: 0.6779 - val_acc: 0.8619
Epoch 2/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.0440 - acc: 0.7923 - val_loss: 1.6138 - val_acc: 0.8619
Epoch 3/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1744 - acc: 0.7923 - val_loss: 1.6138 - val_acc: 0.8619
Epoch 4/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1744 - acc: 0.7923 - val_loss: 1.6138 - val_acc: 0.8619
Epoch 5/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1742 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
Epoch 6/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
Epoch 7/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
Epoch 8/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619

Epoch 00008: ReduceLROnPlateau reducing learning rate to 9.999999747378752e-06.
Epoch 9/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
Epoch 10/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
Epoch 11/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
Epoch 12/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
Epoch 13/100
1815/1815 [==============================] - 70s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
Epoch 14/100
1815/1815 [==============================] - 70s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
Epoch 15/100
1815/1815 [==============================] - 70s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
.....
Epoch 00029: ReduceLROnPlateau reducing learning rate to 1.0000000116860975e-08.
Epoch 30/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
Epoch 31/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
Epoch 32/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
Epoch 33/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
Epoch 34/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
Epoch 35/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
Epoch 36/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619

Epoch 00036: ReduceLROnPlateau reducing learning rate to 9.999999939225292e-10.
Epoch 37/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
Epoch 38/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
Epoch 39/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
Epoch 40/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
Epoch 41/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
Epoch 42/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
Epoch 43/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619

Epoch 00043: ReduceLROnPlateau reducing learning rate to 9.999999717180686e-11.
Epoch 44/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
Epoch 45/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
Epoch 46/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
Epoch 47/100
1815/1815 [==============================] - 71s 39ms/step - loss: 2.1737 - acc: 0.7923 - val_loss: 1.6127 - val_acc: 0.8619
and finally, the evaluation results are very bad:
Area under the ROC curve: 0.5

Area under Precision-Recall curve: 0.6370608990009015

Confusion matrix: Custom threshold (for positive) of 0.5
[[24737000 0]
[ 9341720 0]]
Global Accuracy: 0.7258782019981971
Specificity: 1.0
Sensitivity: 0.0
Precision: 0

Jaccard similarity score: 0.0

F1 score (F-measure): 0.0

So strange！

Hi! Thank you for publishing your code.
When I run prepare_datasets_DRIVE.py in the DRIVE dataset, I found that the test does not contain label, but your program requires the input variable "groundTruth_imgs_test". The test I downloaded contains only images and mask.

imgs max: 0.0
imgs min: 0.0
Traceback (most recent call last):
File "prepare_datasets_DRIVE.py", line 82, in
imgs_train, groundTruth_train, border_masks_train = get_datasets(original_imgs_train,groundTruth_imgs_train,borderMasks_imgs_train,"train")
File "prepare_datasets_DRIVE.py", line 67, in get_datasets
assert(np.max(groundTruth)==255 and np.max(border_masks)==255)
AssertionError
kindly tell me how to resolve this? just tried to run prepare_datasets_DRIVE.py

Thank you for your sharing, and i still confused about the differences between the two coding scheme of bi-directional ConvLSTM .
One is what you discribed before:

LSTM_f = layers.ConvLSTM2D(filters = 128, kernel_size=(3, 3), padding='same', return_sequences = False, go_backwards = False, kernel_initializer = 'he_normal')(X)

LSTM_b = layers.ConvLSTM2D(filters = 128, kernel_size=(3, 3), padding='same', return_sequences = False, go_backwards = True, kernel_initializer = 'he_normal')(X)

Bi_rep = layers.Add()([LSTM_f, LSTM_b])

The other is using the Bidirectional wrapper:

Bidirectional(ConvLSTM2D(filters = 128, kernel_size=(3, 3), padding='same', kernel_initializer='he_normal', return_sequences=True))(X)''.

Could you tell if the second statement is correct, and if there are some differences between them? Than you for your reply！

I believe using go_backwards=True does not make a ConvLSTM bidirectional. It just makes the input reversed. Instead, there should be a Bidirectional layer wrapping around the ConvLSTM to do the job. This way there would be two separate outputs for forward and backward cells. I'm not sure if those should be concatenated or averaged based on the model proposed in the paper. For the first option we have to make the number output channels half, so that the dimensions stay constant.

Hello, thanks for this nice repo.

I was wondering, can I access a pretrained model just to make inference on new images?

From reading the README it seems that I'll need to run the training pipeline first before being able to infer on new images.

I think it would be useful for a lot of people to directly access pretrained weights, would you consider sharing them?

Thanks!

Hello, sir! I hope you are doing well.
While training the model for the skin data set i came across this error:

TypeError: ('Keyword argument not understood:', 'input')

I am training on google colab; the error refers to:

Your help in this regard would be highly appreciated.
Looking forward for your kind response,
Asif Ahmad

Hi!
First of all, thank you very much for sharing your code.

Coud you tell me what is the difference between BCDU_net_D1 and BCDU_net_D3 networks?
I've tried to use BCDU_net_D3 but I have the next error:
OOM when allocating tensor with shape[2,5,256,256,64]

Could you help me?
Thank you very much!

It tried implementing this repo.
For scoring the test dataset, it is taking around 90 mins.

Attached is the screenshot.

Can you please help in order to identify the issue ?

Regards,
Gaurav

Dear rezazad68 , can you tell the tensorflow and keras version? Thanks.