add skeletonization code #65
Conversation
Codecov ReportAttention: Patch coverage is
❗ Your organization needs to install the Codecov GitHub app to enable full functionality. Additional details and impacted files@@ Coverage Diff @@
## main #65 +/- ##
======================================
Coverage 0.00% 0.00%
======================================
Files 40 45 +5
Lines 1411 1545 +134
======================================
- Misses 1411 1545 +134 ☔ View full report in Codecov by Sentry. |
| @@ -0,0 +1,77 @@ | |||
| import numpy as np | |||
There was a problem hiding this comment.
Do we need this function? Could be replaced by torch Gaussian filtering, I think? This should be even faster when running on GPU.
| from scipy.fftpack import fftn, fftshift, ifftn | ||
|
|
||
|
|
||
| def angauss(T: np.ndarray, std: float, r: float) -> np.ndarray: |
There was a problem hiding this comment.
Also, this is a function from Antonio's PySeg. Please make sure to add a note to all of these functions that they come from Antonio's repo :)
| Nx, Ny, Nz = T.shape | ||
|
|
||
| # Initialize arrays for forward and backward differences | ||
| Idp = np.zeros((Nx, Ny, Nz), dtype="float64") |
There was a problem hiding this comment.
Can we make these variable names a bit more descriptive? Would make it easier to read the function
| from skimage.util import img_as_float32 | ||
|
|
||
|
|
||
| def read_nifti(nifti_file: str) -> np.ndarray: |
There was a problem hiding this comment.
We already have all of these functions in
https://github.com/teamtomo/membrain-seg/blob/main/src/membrain_seg/segmentation/dataloading/data_utils.py
Please use these instead of defining them again here :)
| This command runs the skeletonization process from the command line. | ||
| """ | ||
| # labels can be .nii files or mrc files | ||
| if label_path.endswith(".nii") or label_path.endswith(".nii.gz"): |
There was a problem hiding this comment.
I think we can neglect the case of nifti files here. Let's just keep the mrc files.
Also, let's not check for file endings here because there are multiple valid file extensions that work in a similar way and can be processed out-of-the-box by the "load_tomogram" function (e.g. .mrc, .rec, .em, ...).
|
|
||
| # Segmentation consists of 0, 1 and 2 respectively | ||
| # for background, mask and labels to be ignored. | ||
| labels = (segmentation == 1) * 1.0 |
There was a problem hiding this comment.
normally, we will not have the case that we have ignore labels (that's just for the training patches).
So maybe we can make this criterion (segmentation > 0) * 1.0
This will account also for multiple instance values
| hessianZZ = angauss(hessianZZ, std, 1) | ||
| hessianXY = angauss(hessianXY, std, 1) | ||
| hessianXZ = angauss(hessianXZ, std, 1) | ||
| hessianYZ = angauss(hessianYZ, std, 1) |
There was a problem hiding this comment.
Could you please factor out the hessian computation into a separate function?
I.e. a function that computes the gradients, the hessian, and then smoothes the hessian?
| print("Genration of skeleton based on non-maximum suppression algorithm.") | ||
| first_eigenvalue = ndimage.gaussian_filter(first_eigenvalue, sigma=1) | ||
| first_eigenvalue = np.abs(first_eigenvalue) | ||
| Skeleton = nonmaxsup( |
There was a problem hiding this comment.
capitalized variable names indicate classes, so better call it "skeleton" instead of "Skeleton"
| ) | ||
|
|
||
| # Save the skeleton | ||
| write_nifti(save_path, Skeleton) |
There was a problem hiding this comment.
default output should be an .mrc file. Please use
|
Thanks a lot for the updates @Hanyi11 |
No description provided.