[statmlben]

Related to #1172.

Thank you for suggesting an example notebook as a starting point. This PR adds a self-contained tutorial demonstrating RankSEG foreground decoding with TIAToolbox's MicroNet pipeline.

Summary

• Add examples/13-rankseg-postprocessing.ipynb and an entry in examples/README.md.
• Compare argmax and RankSEG using the same probabilities from the pretrained micronet-consep model on Kumar test images.
• Apply MicroNet's instance postprocessor to both decoded masks and evaluate full-image foreground Dice.
• Include an embedded visual comparison, saved execution outputs, and a reproducible evaluation of all 14 test images.

RankSEG is installed as an optional package within the notebook. Model weights and data are downloaded when needed.

Example results

Dice is computed per image and then averaged with equal weight.

Evaluation	Images	Argmax Dice	RankSEG Dice	Gain	Images improved
CoNSeP-trained MicroNet → Kumar test set	14	61.26%	64.94%	+3.68 percentage points	14/14

The illustrated case is explicitly identified as a favorable example selected after reviewing the cohort. Results describe foreground overlap on this small cross-dataset cohort; instance-separation quality requires separate evaluation.

Validation

• Executed the notebook, including the full 14-image evaluation, on an NVIDIA RTX 3090.
• Used Python 3.12, TIAToolbox 2.1.3, and RankSEG 0.0.8 with safe screening enabled.
• Verified that the argmax adapter reproduces the official engine's instance output.
• Passed pre-commit checks for the changed files.
• Colab execution has not been separately verified.

I would appreciate feedback on the notebook's structure and its use of the TIAToolbox 2.x inference API.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB