davlars / ad-skull-reconstruction Goto Github PK

Currently running plot_data on a 2d data set gives an error. Perhaps we should also support 2d?

save_data does not seem to exist in adutils. Instead I find save_image, which seems to do the job. Need to update documentation accordingly (see below).

Save data

To save data in a format that the clinical can review (typically nifti), use the adutils.save_data utility, with x being your reconstruction

fileName = /my/path/myFile
adutils.save_data(x, fileName, as_nii=True, as_npy=True)

Some examples currently have very large windows, these should be changed to something relevant, like [0.018, 0.022]

Currently, all of the "parameter search" functions need to calculate the FoM themselves. We should have a standard library of FoM available so that people can re-use them.

see title

I think that this commit in odl causes crash in the pickled geometries corresponding to this package. The pickled geometries do not have a self.__stride = None in the init so for example power_method_opnorm does not work.

(I am not sure what stride does, but the crash occurred in power_method_opnorm for me)

Basically this makes the load_data_from_nas script fail, at least on windows. This is because Reference and Users are on different drives.

I get the following error:

  File "<ipython-input-1-460109d7ac57>", line 1, in <module>
    runfile('E:/Github/ad-skull-reconstruction/FBP_reco_skullCT_2D.py', wdir='E:/Github/ad-skull-reconstruction')

  File "C:\Anaconda\envs\tensorflow\lib\site-packages\spyder\utils\site\sitecustomize.py", line 866, in runfile
    execfile(filename, namespace)

  File "C:\Anaconda\envs\tensorflow\lib\site-packages\spyder\utils\site\sitecustomize.py", line 102, in execfile
    exec(compile(f.read(), filename, 'exec'), namespace)

  File "E:/Github/ad-skull-reconstruction/FBP_reco_skullCT_2D.py", line 27, in <module>
    phantom = reco_space.element(adutils.get_phantom(use_2D=True))

  File "E:\Github\ad-skull-reconstruction\adutils.py", line 195, in get_phantom
    nii = nib.load(path+phantomName)

  File "C:\Anaconda\envs\tensorflow\lib\site-packages\nibabel\loadsave.py", line 40, in load
    raise FileNotFoundError("No such file: '%s'" % filename)

FileNotFoundError: No such file: '/lcrnas/Reference/CT/GPUMCI simulations/code/AD_GPUMCI/phantoms/70100644Phantom_labelled_no_bed.nii'

This is since we have a hardcoded path to the NAS in get_phantom

There seems to be a bug in how the Tam-Danielsson window is used in the adutils. Currently the code in the adutils reads (see lines 115-119)

if use_window:
    window = odl.tomo.tam_danielson_window(ray_trafo,
                                           smoothing_width=0.05,
                                           n_half_rot=3)
    ray_trafo = window * ray_trafo

which will cause ray_trafo to not be a RayTransform anymore. Thus the standard fbp in odl does not work. Minimal working example to reproduce the crash:

import numpy as np
import adutils

# Discretization
reco_space = adutils.get_discretization()

# Forward operator (in the form of a broadcast operator)
A = adutils.get_ray_trafo(reco_space, use_window=True)

# Define fbp
fbp = adutils.get_fbp(A)

This can be compared to how the Tam-Danielsson window is applied in the example filtered_backprojection_helical_3d in odl.

edit: @adler-j improved typesetting #

The current initial guess has Gibbs artifacts, we should likely find one that is more smooth.

We should add a wavelet regularized example, similar to this one.

The pickled geometries need to be updated to reflect this change.

The line label = np.flipud(label) is redundant and causes the phantom to have the wrong orientation

As mentioned in #19, we should only copy if the data has been modified. See e.g.

http://stackoverflow.com/questions/2266234/can-we-do-a-smart-copy-in-python

The current example in the readme only works properly on windows, a short example on linux usage would probably be popular.

Currently the scripts save to disk with a hard-coded path. We should remove this so that we dont get colliding writes to the NAS.

I have problem with installing requirements. What is going wrong? Anyone else has the same issue?

I write the following command:

:~/ad-skull-reconstruction$ pip install -r requirements.txt

I get the error message:

Collecting odl>=0.6.1 (from -r requirements.txt (line 1))
  Could not find a version that satisfies the requirement odl>=0.6.1 (from -r requirements.txt (line 1)) (from versions: 0.2.2, 0.2.3, 0.3.0, 0.3.1, 0.4.0, 0.5.0, 0.5.1, 0.5.2, 0.5.3, 0.5.3.post0, 0.5.3.post1, 0.5.3.post2, 0.6.0)
No matching distribution found for odl>=0.6.1 (from -r requirements.txt (line 1))

When connect to a remote computer via SSH, start an IPython session and type import adutils (which is properly installed), I get the error message

qt.qpa.screen: QXcbConnection: Could not connect to display 
Could not connect to any X display.

The following code

import odl
import adutils

# Discretization
reco_space = adutils.get_discretization(use_2D=True)

# Forward operator (in the form of a broadcast operator)
A = adutils.get_ray_trafo(reco_space, use_2D=True)

gives this error

Loading geometry
Traceback (most recent call last):

  File "<ipython-input-1-e4354a8993cf>", line 8, in <module>
    A = adutils.get_ray_trafo(reco_space, use_2D=True)

  File "/home/aringh/git/ad-skull-reconstruction/adutils.py", line 86, in get_ray_trafo
    geom = pickle.load(f, encoding='latin1')

ImportError: No module named 'odl.tomo.geometry.fanbeam'

I guess somethings has changed in odl, but I do not know exactly what and when.

This is related to the discussion in #35, and the workaround there works for this as well.

You get the following error when you try to load the geometries

ValueError: unsupported pickle protocol: 3

To solve this, we need to regenerate these with protocol version 2. I.e. someone with python3 needs to do:

pickle.dump(pickle.load(file), file, 2)

on all the geometry files.

The __pycache__ folder has been added to .gitignore but when I did this, i forgot to remove it. We should remove it from the repo.

In several files we have stuff using saveCont that looks like

    niter = [int(i) for i in np.arange(5, 101, 5)]
    for iterations in niter:
        odl.solvers.conjugate_gradient_normal(A, x, rhs, niter=stepiter,
                                              callback=callbackPrintIter)
        # Store to disk

I assume this is intedended to save iterates nr 5, 10, 15, 20, etc, but since we restart the algorithm every 5:th iterate, the so called "10th" iteration is not the 10th iteration if we let conjugate_gradient_normal run its course. This is likely even more evident for the more advanced solvers that "build momentum" towards the optimum.

My suggestion is either to port CallbackSaveToDisk, or wait until it is merged into master and use it instead. This would also make the code much easier to read (simply append a CallbackSaveToDisk to the callback if you want to save results).

The package is becoming crowded, so we should create a folder structure. Perhaps start by adding two folders "reconstruction" and "reconstruction2d"? Naming suggestions are welcome ofc.

Add a saved FBP-reconstruction such that it can be loaded directly (without having to re-run FBP all the time)