jmarcio
commented
2 months ago Hello,
I can't try your example now.
You can convert data between Smil and Numpy using
- getNumpyArray and
- fromNumpyArray
See https://smil.cmm.minesparis.psl.eu/doc/p660.html
Also, watershed algorithms aren't surely not the same in Smil and Scikit Image. Try with a simpler image.
Regards
José-Marcio
On 8/27/24 20:17, Laurids Stockert wrote:
Hello again,
When using the Python wrapper for the watershed algorithm, I would like to obtain the labels as an npy file. Unfortunately, it is not intuitive to me how to transform data between the smil Image datatype and numpy arrays, as well as how to save the labels from the watershed algorithm after it is finished. Below, I have attached an MWE. You may load your own 3d image:
import numpy as np import skimage.io from skimage.morphology import local_minima from scipy.ndimage import label import smilPython as sp import os
|# Convert array to 8-bit unsigned integer format def convert_to_numpy_uint8(data): """Normalize and convert data to 8-bit unsigned integer format.""" data = data.astype(np.float64) / data.max() # Normalize to [0, 1] data = 255 data # Scale to [0, 255] return data.astype(np.uint8) # Convert array to 16-bit unsigned integer format def convert_to_numpy_uint16(data): """Normalize and convert data to 16-bit unsigned integer format.""" data = data.astype(np.float64) / data.max() # Normalize to [0, 1] data = 65535 data # Scale to [0, 65535] return data.astype(np.uint16) def process_image(prediction, output_directory, output_filename): """Process the image and save the result with watershed segmentation.""" # Normalize and invert the prediction inverted_inference = 1.0 - (prediction
- prediction.min()) / (prediction.max() - prediction.min()) # Detect local minima in the inverted inference local_minima_mask = local_minima(inverted_inference) foreground_mask = inverted_inference > 0.4 # Assuming a threshold value of 0.4 for the foreground adjusted_minima = local_minima_mask * foreground_mask # Label the local minima labeled_minima, num_features = label(adjusted_minima) # Convert arrays to the appropriate format for SMIL processing inverted_inference_uint8 = convert_to_numpy_uint8(inverted_inference) labeled_minima_uint16 = convert_to_numpy_uint16(labeled_minima) # Create SMIL Image objects smil_inverted_inference = sp.Image(inverted_inference_uint8) smil_labeled_minima = sp.Image(labeled_minima_uint16) # Compute gradient smil_gradient = sp.Image(inverted_inference_uint8) sp.gradient(smil_inverted_inference, smil_gradient) # Create watershed segmentation smil_watershed = sp.Image(inverted_inference_uint8) smil_basins_out = sp.Image(inverted_inference_uint8) sp.watershed(smil_gradient, smil_labeled_minima, smil_watershed, smil_basins_out) # Retrieve the watershed result watershed_result = smil_basins_out.getNumpyArray() # Save the result to file if not os.path.exists(output_directory): os.makedirs(output_directory) output_filepath = os.path.join(output_directory, f"{output_filename}.npy") np.save(output_filepath, watershed_result) # Print results print("Maximum value in watershed result:", watershed_result.max()) if name == "main": # Load the image image_path = "/path/to/your/image.tif" img = skimage.io.imread(image_path) # Define output directory and filename output_dir = "/path/to/output/directory" output_file = "segmentation_result" # Process the image process_image(img, output_dir, output_file) |
When I do this for my image, I get a maximum label of 65,535. However, I find this very unlikely, as the number of labeled instances I get through scipy.ndimage.label is around 40,000 (the watershed algorithm by scikit image also only gives me around 40,000 instances). How should I correctly save this? Also, as you can see, choosing UINT16 as a datatype here can cause problems. If I had only a few thousand more instances in my image, the labels wouldn't work anymore.
— Reply to this email directly, view it on GitHub https://github.com/MinesParis-MorphoMath/smil/issues/3, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABQE4AG5YS2OOQJUHJ2RPCLZTS7FDAVCNFSM6AAAAABNGVVOPOVHI2DSMVQWIX3LMV43ASLTON2WKOZSGQ4TAMBUG44DAOA. You are receiving this because you are subscribed to this thread.Message ID: @.***>
--
postnubilaphoebus
Hello again,
When using the Python wrapper for the watershed algorithm, I would like to obtain the labels as an npy file. Unfortunately, it is not intuitive to me how to transform data between the smil Image datatype and numpy arrays, as well as how to save the labels from the watershed algorithm after it is finished. Below, I have attached an MWE. You may load your own 3d image:
When I do this for my image, I get a maximum label of 65,535. However, I find this very unlikely, as the number of labeled instances I get through scipy.ndimage.label is around 40,000 (the watershed algorithm by scikit image also only gives me around 40,000 instances). How should I correctly save this? Also, as you can see, choosing UINT16 as a datatype here can cause problems. If I had only a few thousand more instances in my image, the labels wouldn't work anymore.
Thank you again for providing this library, and I look forward to your reply.