The vertebrae_segmentation should input lardmark.csv and valid_lardmark.csv, but the vertebrae_localization has no output form this two files. So how to produce these two files when I predict a new data?

Very nice project and tools, Thanks for sharing! I am wondering how to improve the inference speed. We make a test with 40 3D images on v100 GPU(32GB), it spend 56 seconds per image which is slow for us. Secondly, We want to decrease the model size to use less momory and make inference faster without too much performance drop. What do you think about it?

For data without labels, I find it need possible_successors.pickle and units_distances.pickle in main_vertebrae_localization.py. What is its content and how can i get it? I tried to use pickle folder, but it didn't work. Can you tell me how I can get it?
Thanks!

Hi @christianpayer,
Thank you so much for your source code and your instruction in the Verse.
But I don't quite understand the models files in docker/models.
Could you give me some detail information about that files, please?
Thank you so much.

When training main_ vertebrae_ localization.py, it's very slow,Ican omly train 10000 iters in 24 hours. When some samples are tested, they are stuck. I use one V100 gpu.

After installing tensorflow,I run the code ,I get the simgle import error :

import utils.io.image
ModuleNotFoundError: no module named 'utils.io
and before this error I get the error:

ImportError: No module named 'utils'

After installing the python-utils,the 'utils' work but 'utils.io' doesn't work

My environment: linux ubuntu 16.04 + Tensorflow 1.3.0 + Python 3.6.2

Any help is highly appreciated.

I think many things regarding the dataset have been changed (filenames, orientation,..etc), check the updated dataset here: https://github.com/anjany/verse

I will try to add a PR to address the new changes.

Based on these Slicer Wiki and ITK wiki I think the direction matrix used in the code can also be called LPS orientation, not just RAI, it is matter of using "to" or "from" concepts to represents the letter:

     # LPS: Right to (L)eft, Antiror to (P)ostiror, Inferor to (S)superior
     # RAI: (R)ight to Left, (A)ntiror to Postiror, (I)nferor to Superior
      m = itk.GetMatrixFromArray(np.array([[1, 0, 0], [0, 1, 0], [0, 0, 1]], np.float64))

I think since all output images have the same orientation probably the naming does not matter.

Thanks for sharing the code. I had a problem running the main_vertebrace_localization.py , which generated an empty predictive image with no content in it. For example, the verse004_prediction.mha is empty, no content in it. How can I solve this problem？

10 days ago a new version of SimpleITK (2.0.0) has been released. This is incompatible with TensorFlow 1.14, the base docker container for inference, and results in a Segmentation Fault when running main_spine_localization.py

This codes looks good. however it is too complicated due to the lack of instructions.
After completion of training and inference, I can get the .mha files and bbs.csv
How to verify ( or display) the output results (.mha )?

GL108_CT_ax,0.8339234590530396,8.83392345905304,0.75,96.83392345905304,136.83392345905304,184.75
GL146_CT_ax,3.8125,51.8125,138.875,67.8125,171.8125,242.875
GL195_CT_ax,-12.15625,19.84375,13.25,43.84375,155.84375,141.25
GL216_CT_ax,-3.6572570204734802,52.34274297952652,72.625,108.34274297952652,148.34274297952652,200.625
GL217_CT_ax,-8.64743047952652,31.35256952047348,2.875,95.35256952047348,151.35256952047348,258.875
GL279_CT_ax,9.785285696387291,9.785285696387291,4.5,81.78528569638729,153.7852856963873,180.5
GL348_CT_ax,-14.136589303612709,25.86341069638729,-1.75,73.86341069638729,145.8634106963873,174.25
GL419_CT_ax,-13.15423595905304,18.84576404094696,13.25,74.84576404094696,162.84576404094696,213.25
GL428_CT_ax,4.32711797952652,12.32711797952652,19.375,52.32711797952652,140.32711797952652,107.375
GL492_CT_ax,-14.136589303612709,25.86341069638729,-0.25,89.86341069638729,137.8634106963873,247.75

when i start the training the segmentationscript, the execution stops here.. and don't move on..

when i run main_spine_localization.py file. I get the simgle import error:

InvalidArgumentError: Conv3DBackpropInputOpV2 only supports NDHWC on the CPU.

Any help is highly appreciated.

I would insert intersection_area = max(intersection_area, 0) after line 36, such that two boxes with no overlap have an intersection area of 0.

when i run main_spine_localization.py file. I get the simgle import error:
TypeError: '<=' not supported between instances of 'float' and 'NoneType'
Any help is highly appreciated.

Hello! In the code of "vers2020", “self.optimizer_sigma" doesn't seem to be defined,and When the “ self.learnable_sigma" is true, the code does not work. How to make sigma learnable?

makes sense to me, plz correct me if i misunderstood.

thanks, great work!

Hi,
in The Unet class:
self.prediction = Sequential([Conv3D(num_labels, [1] * 3, name='prediction', kernel_initializer=heatmap_layer_kernel_initializer,
activation=None, data_format=data_format, padding=padding),
Activation(None, dtype='float32', name='prediction')])

Why do you use Activation(None) ? What does it do ? Is this a convolutional operation or a fully connected ?

Thanks !

Hi!

I'm finding that the _seg.png image outputed in inference/main_vertebrae_segementation.py function 'test' is blank.
The file size is 2KB, far smaller than the 50KB input image that is generated just before.

I've checked that the variable prediction_labels_resampled is correct and the _seg.nii.gz output is correct too.

Any ideas on how I could fix this?

Hi,when I train the main_spin_location.py, the loss drop is very small. When training 10000 iters, the loss fluctuates around 800.

Hi, thanks for your open-source toolkit and project. It's very helpful and attractive.
I'm trying to train models of this project. But when I ran the first training file, I met some errors.

This is the information of errors.
The version of Tensorflow is 1.14 and OS is Ubuntu.

Could you tell me how can I solve this problem?
Thank you very much!

Thanks for sharing the code. I had a problem running the main_vertebrace_localization.py , which generated an empty predictive image with no content in it.The error contents is as follows:
Traceback (most recent call last):
File "d:/MedicalDataAugmentationTool/bin/main_vertebrae_localization.py", line 269, in test
curr_landmarks = spine_postprocessing.solve_local_heatmap_maxima(local_maxima_landmarks)
File "D:\MedicalDataAugmentationTool\utils\landmark\spine_postprocessing_graph.py", line 139, in solve_local_heatmap_maxima
shortest_path = nx.shortest_path(G, 's', 't', 'weight', method='bellman-ford')
File "D:\ana\envs\verse\lib\site-packages\networkx\algorithms\shortest_paths\generic.py", line 164, in shortest_path
paths = nx.bellman_ford_path(G, source, target, weight)
File "D:\ana\envs\verse\lib\site-packages\networkx\algorithms\shortest_paths\weighted.py", line 1385, in bellman_ford_path
length, path = single_source_bellman_ford(G, source, target=target, weight=weight)
File "D:\ana\envs\verse\lib\site-packages\networkx\algorithms\shortest_paths\weighted.py", line 1612, in single_source_bellman_ford
dist = _bellman_ford(G, [source], weight, paths=paths, target=target)
File "D:\ana\envs\verse\lib\site-packages\networkx\algorithms\shortest_paths\weighted.py", line 1268, in _bellman_ford
raise nx.NodeNotFound(f"Source {s} not in G")
How can I solve it?

i want to add more dataset to improve performance but i have issue with creating landmarks labels for the dataset, any software that can help with labels?

Hi,

when I try the link https://hub.docker.com/repository/docker/christianpayer/verse20 that is mentioned in verse2020/docker it seems I do not have access to the image. It asks me to sign in and when I sign in I get a 404. Is this not available any more or did you configure it wrong? Or am I missing something?

Thanks,

Eren Yilmaz

Thank you very much for your open source code which helps me a lot. I have a question about the landmark.csv which is a littile different from the official JSON file. It much confuse me about the X coordinate value and how to generate my own landmark.csv.

I can't understand why to Dataset preprocessing，and what is RAI ？can you explain for me?
TK u

It about the size of GLxxx file. I tried to get the 2D - central image from the 3D image. Here is the result of GL 003.

Do y know why the size of these files is not correct?

When training main_spine_localization.py, loss_net = nan. I had check all parameter,but all this.
Do you know why? thank you

09:37:40: train iter: 0 loss_net: 0.3521 norm: 4.5986 norm_average: 9.9460 seconds: 11.332
09:38:08: train iter: 100 loss_net: nan norm: nan norm_average: 9.9460 seconds: 28.198
09:38:37: train iter: 200 loss_net: nan norm: nan norm_average: 9.9460 seconds: 29.554
09:39:06: train iter: 300 loss_net: nan norm: nan norm_average: 9.9460 seconds: 28.922
09:39:40: train iter: 400 loss_net: nan norm: nan norm_average: 9.9460 seconds: 33.664
09:40:12: train iter: 500 loss_net: nan norm: nan norm_average: 9.9460 seconds: 32.590