borda / birl Goto Github PK

evaluate already finished experiment and recompute all required metrics from the input information

• refactor benchmark class
• ensamble evaluation script

configure CodeShip Docker for py2 and py3

transfer registration methods once used in previous registration benchmarks

• ASSAR affine
• ASSAR bSpline

transfer registration methods once used in previous registration benchmarks

• jlSegReg affine
• jlSegReg elastic

transfer registration methods once used in previous registration benchmarks

• RNiftyReg linear
• RNiftyReg nonlinear

transfer registration methods once used in previous registration benchmarks

• TrakEM affine
• TrakEM bSpline

In case we have registration method that does not provide transformed landmarks we want to detect then by our-self. We take the list of landmarks in the moving image and try to detest those landmarks in the registered image. Assumption: the deformation between these two images is not large.

• find reasonable descriptors
• match landmarks
• robustness: run in multiscale

Python Image Registration Toolkit

Traceback (most recent call last):
  File "bm_dataset/create_real_synth_dataset.py", line 329, in <module>
    main(arg_params)
  File "bm_dataset/create_real_synth_dataset.py", line 313, in main
    for _ in mproc_pool.imap_unordered(wrapper_deform_export, range(params['nb_samples'])):
  File "/usr/lib/python3.7/multiprocessing/pool.py", line 748, in next
    raise value
  File "/usr/lib/python3.7/multiprocessing/pool.py", line 121, in worker
    result = (True, func(*args, **kwds))
  File "bm_dataset/create_real_synth_dataset.py", line 263, in perform_deform_export
    image_out, points_out = deform_image_landmarks(image_out, points, max_deform)
  File "bm_dataset/create_real_synth_dataset.py", line 163, in deform_image_landmarks
    zip(x.ravel(), y.ravel()), image.reshape(-1, 3), (x + x_deform, y + y_deform), method='linear', fill_value=1.
  File "/usr/local/lib/python3.7/dist-packages/scipy/interpolate/ndgriddata.py", line 221, in griddata
    rescale=rescale)
  File "interpnd.pyx", line 245, in scipy.interpolate.interpnd.LinearNDInterpolator.__init__
  File "interpnd.pyx", line 80, in scipy.interpolate.interpnd.NDInterpolatorBase.__init__
  File "interpnd.pyx", line 190, in scipy.interpolate.interpnd._check_init_shape
IndexError: tuple index out of range

Issue

While creating synthesis data by executing bm_dataset/create_real_synth_dataset.py I am facing the above error .I think it might be because of the scipy version issue.
I am using scipy==1.4.1 and also tried with scipy 1.1.10
I am using python version 3.7.13 in google colab
Please guide me and help me to solve issue

Add some extra landmarks validation, more then just distance

• estimate affine transformation for init pair
• compare the affine registration of GT
• compare difference after subtraction of affine transform
• decompose the affine transform to shift, rotation, scale

Asked (reported) to ImageJ: I have a case - register two images together, then warp the source image to target image and in the end generate the raw transformation to be able show some important points in the warped image according their positions in source image. (In this moment I don't race much about quality of the registration.) This perfectly works when both image are the same size but in following case the source image is smaller then the target one.

See an example if correctly transformed landmarks

bad points warping in the warped image

and some more visualization where it should be

• macro I used for registration, warping image and converting raw transformation

transfer registration methods once used in previous registration benchmarks

Hello @Borda
I have followed your advice in regards to running DROP2 through BIRL and I have a couple of questions. The code generates these images and I'm not sure which one is aligned to which one? And I believe the output.jpeg file is the output from DROP2 before applying the transformations onto the RGB?

I tried finding the code for the transformation on the RGB image and found the following, which appears to apply a shift on the landmarks. Can you point me out to the specific code performing the transformation on the original RGB image given output_field_x.nii.gz and output_field_y.nii.gz?
shift = self.extract_landmarks_shift_from_nifty(path_deform_x, path_deform_y, lnds_)

Finally, if the application of the transformation on the RGB image ends up being difficult to resolve on my end given the info you provide above (I hope not, but I'm new to image analysis so it is possible), can I just use the BIRL functions to perform the registration and transformations on the image, without providing any landmark information?

Thanks

transfer registration methods once used in previous registration benchmarks

• ANTs-affine
• ANTs-elastic

You are referring to the wrong library
While OpenSlide is an image library written in C, OpenSlides is a tool to manage big assemblies.

https://github.com/Borda/BIRL/blob/master/bm_dataset/convert_tiff2png.py#L38

🚀 Feature

Create a Docker containing all mentioned SOTA installed and upload to docker hub

Motivation

just out-of-the-box for any experimenting

Pitch

we could use this image even for testing all listed methods

have look at

• pyElastix
• SimpleElastix

prepare evaluation script for ANHIR according to https://anhir.grand-challenge.org/Evaluation

• evaluation script
• normalize exec. time
• match related landmarks
• export JSON results
• docker image

write a short tutorial how to annotate the landmarks using ImageJ and add set of tools for validation of this landmarks

• visualize the landmarks change between each two images
• merge annotation from myltiple users (the same file names in different folders)
• scaling images and landmarks into specific scale
• compute the internal landmarks distances for each image and assume that this graph should not change much among images

create experiment with no registration method which would serve as test case
this experiment would serve as baseline for all experiments, just override perform method

• parsing inputs
• loading cover csv file
• run sequential experiments
• run experiments in parallel
• evaluate experiment
• summarize results
• visualize: reference - registered image
• visualize: moving - registered image

transfer registration methods once used in previous registration benchmarks

• RVSS affine
• RVSS elastic

Description

there is a certain issue with converting SVS image using convert_tiff2png
the images mooks like shifted and incomplete...

create a ipython notebook with loading finished benchmarks and do some simple comparison

• simple bar chart or error bar for each comaprison subject (like time, inaccuracy, ...)
• create n-shape diagram with edges as comparison subject (like time, inaccuracy, ...)

transfer registration methods once used in previous registration benchmarks

• elastix-OpenCV bspline
• elastix-OpenCV affine
• openCV-SURF

transfer registration methods once used in previous registration benchmarks

Implement general experiment and inherit from this class all other experimented regist. methods.

For evaluation use a weight which would assign a high weight to "alone" landmarks compare to a group of landmarks would have a smaller weight. This will make the evaluation more uniform for any landmark placement. For landmarks ompute the minimal distance to the nearest landmark and normalize this distance by mean of this distances.

Create a parent class for all experiment that in its base evaluates statistic on unregistered image