What's Changed

• v1.0.1: 2d + 3d code examples by @lmanan in #6
• v0.2.3 by @lmanan in #7
• Automatic calculation of crop size by @lmanan in #13
• v0.2.5 - tag (a) by @lmanan in #19
• V0.2.5 - tag (b) by @lmanan in #20
• V0.2.5 - tag (c) by @lmanan in #21
• V0.2.5 - tag (d) by @lmanan in #22
• Update train.py by @ajinkya-kulkarni in #29
• Update utils.py by @ajinkya-kulkarni in #28

New Contributors

• @ajinkya-kulkarni made their first contribution in #29

Full Changelog: v0.2.0...v0.2.5

This release was used to compute numbers for the MIDL publication and is stable.

• The normalization of the image intensities was done by dividing pixel intensities by 255 (for 8-bit images) and 65535 (for unsigned 16-bit images). While this normalization strategy lead to a faster training, it lead to a sometimes, poorer OOD performance. In the future releases, the default will be set to min-max-percentile (takes model longer to reach the same val IoU but leads to a better inference performance).

A minor update since release v0.2.2. This includes:

• Add display_zslice parameter and save_checkpoint_frequency parameter to configs dictionary here

1. Support for visualization for setups when virtual_batch_multiplier > 1 is still missing.
2. Also hardcoded install version of tifffile in setup.py here because latest version currently (2021.6.14) generates a warning message with imsave command while generating crops with bbbc010-2012 dataset. Will relax this version specification in release v0.2.4

TODOs include:

1. Plan to update pytorch version to 1.9.0 in release v0.2.4 (currently pytorch version used is 1.1.0)
2. Plan to add tile and stitch capability in release v0.2.4 for handling in large 2d and 3d images during inference
3. Plan to add a parameter max_crops_per_image in release v0.2.4 to set an optional upper bound on number of crops extracted from each image
4. Plan to save all instance crops and center crops as RLE files in release v0.2.4
5. Plan to add an optional mask parameter during training which ignores loss computation from certain regions of the image in release v0.2.4
6. Plan to deal with bug while evaluating var_loss and to have crops of desired size by additional padding.
7. Plan to include support for more classes.
8. Normalization for 3d ==> (0,1, 2)
9. Make normalization as default option for better extensibility
10. Parallelize operations like cropping
11. Eliminate the specification of grid size in notebooks -set to some default value
12. Simplify notebooks further
13. Make colab versions of the notebooks
14. Test center=learn capability for learning the center freely
15. Add the ILP formulation for stitching 2d instance predictions
16. Add the code for converting predictions from 2d model on xy, yz and xz slices to generate a 3D instance segmentation
17. Add more examples from medical image datasets
18. Add threejs visualizations of the instance segmentations. Explain how to generate these meshes, smoothen them and import them with threejs script.
19. Padding with reflection instead of constant mode
20. Include cluster_with_seeds in case nuclei or cell detections are additionally available

• Add all 3d example notebooks
• Pad images with average background intensity instead of 0

Major changes:

• Add 3d example notebooks for two datasets
• Correct min_object_size (evaluated now from looking at the train and validation masks)
• Save tif images with datatype np.uint16 (in the prediction notebooks )
• Provide support in case evaluation GT images are not available (during prediction)

Some things which are still incorrect in v0.2.0:

• n_y should be set to n_x for equal pixel/voxel sizes in y and x dimension. This is fixed in v0.2.1
• anisotropy_factor is wrongly calculated for the 3d notebooks (it was calculated as the reciprocal). This is fixed in v0.2.1
• train_size was set to 600 for the bbbc012-2010 dataset. This is raised to 1200 in v0.2.1

• Initial functional 2d code (min_object_size was hard coded to 36 and will be updated in later iterations)
• Assets include:
  • 2d images and GT instance annotations
  • 3d images and GT instance annotations
  • fully trained models (*demo.tar) (models trained from scratch up till 200 iterations)
  • glasbey-like colormap (cmap_60.npy)

Data is released under CC-BY 4.0 license (see license).