3DSlicer exploration - interpolation in 3DSlicer

[maxradx]

Manual correction script is mainly designed for using with fsleyes, a open-source software that easily displays medical images in nifti format. However, navigation in fsleyes is subefficient from a user-perspective for performing manual segmentations / manual corrections.

3DSlicer exploration has been performed with promising results. However, 3DSlicer has also its disadvantages (e.g. challenges related to python interpreter from external API ....). Related to deep-learning model training, one of its potential disadvantage is the interpolation done in the default 3DSlicer configuration.

The interpolation from 3DSlicer can by deactivated:

• Manually (in the volume extension, by unclicking on interpolate)
• Automatically by creating a python function that disable it each time a volume is loaded (segmentation are also affected since if there is no interpolation, the segmentation masks will not be interpolated either --- tested locally)
• By ToggleButton and/or keyboard shortcurts -- this can be useful for annotators since an interpolated image is usually more efficient for lesion visualisation and estimating its boundaries. Thereby, an efficient way of switching between interpolated images for a given volume and non-interpolated images is essential (or highly suggested) in manual segmentation / manual correction tasks.

This video (visual only) shows an example that can be achieved in using 3DSlicer: this appears more challenging to achieve in fsleyes.

TO DO :

• Release a 3DSlicer module for optimized manual segmentation/manual correction that include the interpolation activation/deactivation function easily managed:
  - [ ] From keyboard shortcut -- needs to be added.
• Discuss with neuropoly team if interpolation can be toggled while performing manual segmentations / manual corrections (in fsleyes) --- an annotator should have the possibility to display easily interpolated images (for better human visualization) and/or non-interpolated (for segmentation masks used in model training).

Questions.

• 1. Is the interpolation in 3DSlicer linear?
• 2. Is it possible to train a model from interpolated images?
• 3. Is there any chance that it would be equal and/or better?
• 4. What are the limitations when training a deep-learning based segmentation model on interpolated images?