MEDOTTER · A UNIFIED SEGMENTATION PLATFORM

Pick a scan.
Get a segmentation.

MedOtter is a research platform for 3D medical image segmentation. It brings model selection, inference, and editable masks into one planned workflow. Public app and SDK builds are still under release validation, with deployment targets spanning laptops and clusters.

View availability → Paper status →
MedOtter multi-organ segmentation on an abdominal CT slice
CT · MULTI-ORGAN
MedOtter breast mass segmentation on an ultrasound image
ULTRASOUND · BREAST MASS
MedOtter surgical instrument segmentation on an endoscopy frame
SURGERY · INSTRUMENTS
ZERO-SHOT Many modalities, one platform CT · ultrasound · endoscopy - same SDK
WHY MEDOTTER

The models exist.
Using them is the hard part.

Hundreds of brilliant models. Hundreds of incompatible repos. Days of setup to segment one scan, and no fair way to compare any of them. MedOtter is the fix: one zoo, one SDK, one benchmark, local by default.

A platform, not
just a model.

MedOtter is being built to unify the fragmented landscape of 3D medical image segmentation into a single, opinionated workflow. Its design connects model selection, inference, and editable masks in a viewer intended for daily research use.

Many modalities. Many tasks. One pipeline. Whether you work with CT, MRI, ultrasound or pathology slides, the same interface is intended to handle it - without gluing together six different toolchains.

Training is part of the planned interface, alongside inference and evaluation. Published models and datasets remain browsable on Hugging Face, while these docs describe the evolving workflow and the leaderboard records results with their evaluation context.

MODEL ZOO 01
106 models, not one monolith.
A curated zoo of segmentation architectures - U-Net to Transformers, Mamba and RWKV - browsable on Hugging Face while public SDK integration remains in preview.
↗
DATASETS 02
128 datasets onboarded.
Public and in-house cohorts cataloged for a common loader interface, with public datasets mirrored on Hugging Face. The count updates automatically.
↗
TASKS 03
Benchmarked by task.
Each segmentation task is evaluated and ranked on the leaderboard. The first task is in review now; the count here stays in sync as new tasks land.
1 in review
TRAINABLE 04
Training workflows in preview.
The training surface and configuration model are documented while public SDK packaging is finalized. Release-ready installation steps will be published with the first supported build.
→
APP 05
A desktop app in preview.
The multi-planar 2D + 3D viewer is undergoing release validation. Public installers will appear on the availability page once they are ready.
→
DEPLOY 06
Local deployment is the target.
The runtime is designed for CPU and Apple MPS, with the same interface scaling to CUDA. Public builds remain gated on release validation.
→
PLATFORM STRUCTURE

One platform, three paths through it.

01 Your scan

CT, MRI, ultrasound or pathology - NIfTI / DICOM in, no conversion homework.

02 Model zoo

106 curated models; the platform picks the right one by modality and task.

03 Local inference

One SDK on CPU, Apple MPS or CUDA - your data never leaves the machine.

04 Editable mask

Opens in the built-in viewer: brush, lasso, propagate, volumetry.

TRAIN
Your data (128 public sets onboarded) Same SDK - one train config Fine-tuned model joins the zoo
get the SDK →
BENCHMARK
Submit a Hugging Face model ID GPU eval worker · hidden test split Ranked on the public leaderboard
see the leaderboard →
ADOPTION

Datasets people actually pull.

Rolling 30-day downloads across every MedOtter cohort on Hugging Face. Hover the line to read any day; the panel ranks today's most-pulled datasets. Updated daily.

281,449downloads · rolling 30-day · updated 2026-07-28
0131K263K394K05-1506-0107-0107-28
Top datasets · 30-day
  • 4D-Lung: 117,278 downloads
  • NSCLC-PleuralEffusion: 12,819 downloads
  • QIN-LungCT-Seg: 9,848 downloads
  • LUNA16: 9,046 downloads
  • HCC-TACE-Seg: 8,067 downloads
  • TCIA_CervicalCancer: 7,799 downloads
  • CT_Lymph_Nodes: 7,548 downloads
  • QIN-PROSTATE: 6,886 downloads
  • TotalSegmentatorMR: 6,366 downloads
  • PI-CAI: 5,427 downloads

Zero-shot segmentation across 6 cases in 4 imaging modalities - one model, no fine-tuning, prompt-free. Drag each divider to compare the raw scan with MedOtter's prediction. Dice scores are dataset means; live model scores are on the leaderboard.

Polyp
endoscopy · KvasirSEG
Dice 0.87
Polyp raw scan
Polyp prediction overlay
INPUT
SEGMENTATION
Skin lesion
dermoscopy · ISIC2018
Dice 0.91
Skin lesion raw scan
Skin lesion prediction overlay
INPUT
SEGMENTATION
Breast mass
ultrasound · BUSI
Dice 0.91
Breast mass raw scan
Breast mass prediction overlay
INPUT
SEGMENTATION
Surgical instruments
endoscopy · Endovis2017
Dice 0.86
Surgical instruments raw scan
Surgical instruments prediction overlay
INPUT
SEGMENTATION
Kidney
ct · KiPA22
Dice 0.58
Kidney raw scan
Kidney prediction overlay
INPUT
SEGMENTATION
Abdominal organs
ct · FLARE22
Dice 0.78
Abdominal organs raw scan
Abdominal organs prediction overlay
INPUT
SEGMENTATION