Engine·Apps·Edge·Agents·Scout·FidOpt

Medical-imaging AI, from the dataset to the browser

KonfAI is a declarative execution engine and the family around it. One configuration trains, predicts and evaluates a method; Scout inspects its results, FidOpt scores it on 311 datasets, and it ships as an app to the command line, 3D Slicer, an agent or a browser tab.

Get started See the whole loop

Code·Docs·Hugging Face·Paper

The whole life of a methodSix stages, one family. A person drives each one, and so can an agent.

Your wayFrom the command line, Python, 3D Slicer or a chat in KonfAI Studio; an agent reaches the same stages through KonfAI MCP.

Data
Your images
ITK images, DICOM, HDF5 and OME-Zarr, read region by region.
./Dataset:mha
Build
KonfAI
One YAML trains, predicts, evaluates and transforms.
Config.ymlkonfai TRAIN
Inspect
Scout
Reads the result like a person at a viewer and boxes what is wrong.
verdict + boxeskonfai-scout inspect
Score
FidOpt
Held-out cases of 311 datasets, the same rules for every method.
the boardfidopt run
not better yet: tune, retrain, score again
Ship
KonfAI Apps
Packaged once, run from the CLI, konfai-apps, 3D Slicer or Hugging Face.
app.jsonkonfai-apps infer
soonAnywhere
KonfAI Edge
The same app in a browser tab, with no server.
<konfai-app>open a page

Measured on 311 datasetsFidOpt scores every method on the same held-out cases, against truth that experts drew. A sample of its corpus:

The boards on fidopt.io ↗

**Pediatric T1 MRI with and without previous ADHD diagnosis**registration

**DeePiCt S. pombe cryo-electron tomograms**segmentation

**Myocardial perfusion SPECT**registration

**Colorectal cancer TMA cores, H&E to CD44**registration

**ISIC 2018, skin lesions outlined on dermoscopy**segmentation

**SNEMI3D sections, known in-plane deformation**registration

**CPTAC-UCEC-Tumor-Annotations**segmentation

**QIN-PROSTATE-Repeatability**registration

**CSXA cervical radiographs**registration

**LCTSC CT and simulated CBCT, known deformation**registration

**HC18, the fetal head as a sonographer measured it**segmentation

**Kvasir-SEG, colonoscopy polyps outlined by hand**segmentation

**Lausanne TOF-MRA aneurysms**segmentation

**BUS-BRA, breast ultrasound tumours outlined by hand**segmentation

**Cardiac CT of growing pigs (A heart for pigs)**registration

**AIND SmartSPIM whole mouse brains, two channels**registration

311datasets, one evaluation for all

3,353test cases, their truth kept on the server

2,960development cases, their truth included

3tasks: registration, segmentation, synthesis

Where it goesMethods that improve in a loop, judged in the open.

An agent that develops a method should be able to ask a trusted service how good it is, and iterate, the way a person submits to a challenge, but in a loop.
FidOpt

Today

Every stage runs

KonfAI trains, predicts and evaluates from one config; Scout reads the results; FidOpt scores methods on 311 datasets; five apps run from the CLI, 3D Slicer or an agent.

Every app in a browser

KonfAI Edge runs the apps in a browser tab with no server, as Web Components another page embeds with one import. Its demo lands on this site.

The direction

The loop, closed

An agent writes a method as a KonfAI app. KonfAI runs it, Scout inspects it, FidOpt judges it on a hidden test, and what wins ships as an app anyone can run.

The family, piece by pieceEach piece stands alone; together they cover the whole loop.

Engine

KonfAI

Patch-native training, complete medical-image inference and a reusable application from one configuration, without giving up the PyTorch and MONAI components you already trust.

Get started → Apps

KonfAI Apps

Five published bundles: TotalSegmentator-KonfAI, MRSegmentator-KonfAI, ImpactSynth, ImpactSeg and IMPACT-Reg. One contract, from a local folder, a Hugging Face repository or an HTTP server; from a one-line CLI, konfai-apps or 3D Slicer.

See the apps → Inspection

KonfAI Scout

Looks at a segmentation, a registration or a synthesis the way a person would at a viewer, and reports where it is wrong, with the views that show it and KonfAI Eval's measurements.

How it reads a result → Studio

KonfAI Studio

A chat web UI over the MCP server. Point it at a dataset and, from the conversation, inspect, train, predict, evaluate, compare and look at the volumes. Your own model endpoint, no API key required.

Open Studio → Agents

KonfAI MCP

Fifty tools, one per object. An agent inspects a dataset, writes and validates a config, launches and follows runs, compares them and packages the result, from Claude Code, Codex or any MCP client.

The fifty tools → coming to this siteEdge

KonfAI Edge

The Rust runtime that runs KonfAI apps in a browser tab, with no server: inference, evaluation and uncertainty beside one viewer, as Web Components another page embeds with one import.

What it is → Benchmark

FidOpt

One open benchmark for biomedical image registration, segmentation and synthesis, every method scored on the same test cases under the same rules. KonfAI MCP reads its catalogue and boards and launches its runs.

fidopt.io ↗ Docs

Documentation

Ten short pages, one per subject: install, configuration, workflows, datasets, apps, agents, Scout, Studio, Edge. The exhaustive reference stays on Read the Docs.

Read the docs → Research

The paper

KonfAI's design and the six MICCAI 2025 challenge results obtained with it, SynthRAD2025, CURVAS, TrackRAD2025 and Panther among them.

Read on arXiv ↗

How it worksOne config, one command per workflow, one app at the end.

# examples/Registration/Config.yml, in part
Trainer:
  Model:
    classpath: registration.registration.VoxelMorph
    VoxelMorph:
      dim: 2
      int_steps: 7
$ pip install "konfai[imaging]"
$ konfai TRAIN -c Config.yml --gpu 0
$ konfai PREDICTION -c Prediction.yml --models Checkpoints/<run>/<checkpoint>.pt --gpu 0
$ konfai EVALUATION -c Evaluation.yml
$ pip install konfai-apps
$ konfai-apps infer VBoussot/TotalSegmentator-KonfAI:total -i ct.nii.gz -o out/

The commands, in order: pip install "konfai[imaging]"; konfai TRAIN -c Config.yml --gpu 0; konfai PREDICTION -c Prediction.yml --models Checkpoints/<run>/<checkpoint>.pt --gpu 0; konfai EVALUATION -c Evaluation.yml; pip install konfai-apps; konfai-apps infer VBoussot/TotalSegmentator-KonfAI:total -i ct.nii.gz -o out/.

At a glanceEvery line opens its page.

The ways in, the same engine behind each

The command linekonfai TRAIN · PREDICTION · EVALUATION · TRANSFORM→ Pythonkonfai.train · konfai.predict_model · konfai.transform→ 3D SlicerSlicerKonfAI · SlicerImpactReg→ Hugging Facekonfai-apps infer VBoussot/<App>→ A conversationKonfAI Studio, your own model endpoint→ An agentKonfAI MCP, from Claude Code, Codex or any client→ A browser tabKonfAI Edge, no server · soon→

Published apps, by task

Segmentation 3TotalSegmentator, MRSegmentator, ImpactSegSynthesis 1ImpactSynthRegistration 1IMPACT-Reg

The docs, one page each

Get started Configuration Workflows Datasets Apps Agents and MCP Scout Studio Edge

4workflowsTRAIN, PREDICTION, EVALUATION, TRANSFORM

5published appssegmentation, synthesis, registration

50MCP toolsone per object, for any MCP client

6MICCAI resultsSynthRAD2025 Tasks 1 and 2, CURVAS PDACVI, TrackRAD2025, Panther, CURVAS

Medical-imaging AI, from the dataset to the browser

The whole life of a methodSix stages, one family. A person drives each one, and so can an agent.

Your images

KonfAI

Scout

FidOpt

KonfAI Apps

KonfAI Edge

Measured on 311 datasetsFidOpt scores every method on the same held-out cases, against truth that experts drew. A sample of its corpus:

Where it goesMethods that improve in a loop, judged in the open.

Every stage runs

Every app in a browser

The loop, closed

The family, piece by pieceEach piece stands alone; together they cover the whole loop.

KonfAI

KonfAI Apps

KonfAI Scout

KonfAI Studio

KonfAI MCP

KonfAI Edge

FidOpt

Documentation

The paper

How it worksOne config, one command per workflow, one app at the end.

At a glanceEvery line opens its page.