Getting started

Introduction

Public package status: the Python distributions and desktop installers are not released yet. These pages document an evolving interface, not currently usable install artifacts. View release status.

What MedOtter is

MedOtter is a medical-image segmentation benchmark and leaderboard platform. The public site reports approved evaluations across radiology and imaging tasks, while a Python interface for catalog, inference, training, and local evaluation is being prepared for release.

The planned Python SDK is the programmatic entry point to that platform: browse the dataset catalog, load data as uniform samples, run inference with SAM2 or a trained checkpoint, and score predictions with the same metrics the leaderboard uses.

The two packages

The development source is organized as two planned packages for different audiences:

Package	Import	Audience	Public status
`medotter`	`import medotter as mo`	Researchers: dataset catalog, data loading, inference, training, evaluation	Release pending
`medotter_bench`	`import medotter_bench`	Leaderboard submitters and the GPU evaluation worker: torch-free scoring, benchmark bundles, prompt-free inference contracts	Release pending

The similarly named PyPI distributions are placeholders/name reservations, not supported MedOtter releases. See Installation status for the current public availability.

After public release, researchers will use the planned medotter facade for catalog, inference, training, and evaluation workflows. The narrower medotter_bench interface is intended for model packaging and scoring contracts. Until supported artifacts exist, the examples in this guide are interface previews; see Installation status.

The dependency-light design

import medotter costs only numpy. The catalog (list_datasets, describe), the Sample and DatasetInfo data classes, and label semantics (LabelSchema) are pure Python and numpy — no torch, no scipy, no network access.

Everything heavier is pulled in lazily, on first use:

mo.load_dataset(...) imports torch and the dataloader package the first time you call it.
mo.load_model(...) and mo.train(...) import torch (and, for SAM2, the model backbone) only when called.
The first call to mo.score(...) or mo.evaluate(...) — with any metrics=, not just hd95 — imports the scoring stack, which pulls in scipy. Accessing mo.eval directly does the same. A catalog-only script that never scores anything never pays for it.

In the development source, this keeps a catalog-only import medotter lightweight without pulling in the full ML stack. The contract still needs to be validated against the first supported package artifact before release.

Where to go next

Reviewing the planned SDK? Start with Installation status, then read the Quick Start as an interface preview.
Want to go deeper on a specific area — the dataset catalog, data loading, inference, evaluation, training, or 3D volume benchmarks? See the Guides section in the sidebar.
Submitting a model to the public leaderboard? See the Leaderboard section in the sidebar for the submission contract and packaging steps.