MONAI

RankSEG is featured in the official MONAI Tutorials as an optional third-party post-processing method. The tutorial shows how to wrap RankSEG as MONAI-style array and dictionary transforms and insert them into a standard decollated inference pipeline.

Important

The tutorial defines RankSEG and RankSEGd wrappers for integration with MONAI workflows. Their inclusion in the official MONAI Tutorials does not mean that RankSEG is built into every MONAI release. Install the rankseg package explicitly and follow the tutorial for the integration shown there.

What the tutorial covers

• loading the public pancreas_ct_dints_segmentation MONAI Bundle;

• running inference on a 3D CT volume from Medical Segmentation Decathlon Task07 Pancreas;

• applying RankSEG and RankSEGd after softmax probabilities are produced;

• comparing RankSEG with AsDiscrete(argmax=True) on the same probabilities;

• preserving MONAI’s channel-first and dictionary-transform conventions.

The model remains frozen: RankSEG replaces only the final probability-to-mask decision and requires no retraining or method-specific fine-tuning.

Run the tutorial

• View the notebook in the official MONAI Tutorials repository

• Open the notebook in Google Colab

Interpreting the comparison

The notebook demonstrates the integration and a paired comparison on a real medical segmentation workflow. RankSEG’s effect depends on the model’s probability quality, the selected metric, and the deployment distribution. Validate both metric changes and post-processing cost on representative data before adopting it in production or reporting a general performance claim.