Documentation

Contents
Quick Start Guide Study Design AI Models Input Requirements Output Format DICOM folders Privacy Vault The Viewer Provenance Workflow Overview
Quick Start Guide
Get your first AI job running on the GPU cluster in under five minutes.
1
Create your account using the Sign-up page. A Guest tier account is created automatically — no payment (details) required.
2
Verify your email after registration. A 6-digit code will be sent to your address. Enter it to activate your account.
3
Submit a job. Navigate to Submit Job, upload a NIfTI image (.nii.gz), a DICOM folder, or another supported format, select an AI model, and click Submit to GPU Cluster.
4
Monitor progress on the Job Queue page. Jobs are picked up in priority order by the GPU cluster, and several workers may process jobs at the same time.
5
Download & preview. When complete, your job appears in My Results. Preview the result in-browser, and/or download the .nii.gz model output.
6
Rate the result (optional). Your ratings shape our public model leaderboard — helping every researcher cut through the noise and pick the AI that actually performs.
Study Design
Organise your imaging into studies, patients and scans — then dispatch a single scan, a whole patient, or an entire cohort to any AI model. Results flow straight back into the study.
1
Create a study on the Study Design panel. Optionally assign the AI models you use most — they become quick-pick chips and pre-filter the model picker. Each study can use a different set of models.
2
Add patients to the study using a pseudonymised code or ID. Patients can be moved between studies, edited, or deleted at any time (icons on each row).
3
Add scans to a patient: upload the NIfTI image (.nii.gz), pick the modality (CT / MRI / PET-CT / Other) and enter the scan date. Scans are stored on the platform, so they persist across sessions and can be re-used.
4
Send to a model. Right-click a scan (or a patient / cohort) for a quick menu, or use the send icons. Pick a model — the estimated credit cost is shown before you confirm. Scans whose modality doesn't match the model are flagged, but can still be sent after a confirmation. You can also open Submit Job → From Study Design to multi-select scans (with Select all) and send them with the full model selector.
5
Track results in place. Each scan shows its jobs' live status. A queued job can be withdrawn (credits refunded) from the status badge; a completed job can be previewed or downloaded directly from the study. Every job also appears as normal in Job Queue and My Results.
AI Models
The platform's model catalog is loaded live from the registry. Filter the table below to find a model by modality, dimensionality, or goal.
The gift icon marks models and tasks that are part of the Nalvera partnership program — every job you run on one of them routes a small portion of the flat processing fee back to the model’s original developers / the open-source community that maintains it. Filter by Partnership program below to see only program models, or read more in the FAQ.
Loading model catalog…
Input Requirements
Ensure your scan meets these specifications before submission.
  • Format: NIfTI .nii.gz, or a DICOM folder (converted automatically on upload). Several other common research imaging formats are supported too — they're converted to NIfTI in your browser before upload, so only the de-identified result is sent to the platform.
  • Encrypted input: a .nvlt export from Nalvera Privacy Vault is also accepted — the platform already decrypts it automatically before validation on Submit Job, Study Design, and the Interactive Studio. Privacy Vault itself is still in development, so you'll be able to produce these encrypted exports once it ships.
  • Maximum file size: 100 MB per submission.
Check the model card on the Submit Job page before sending data — submitting to a mismatched model produces unreliable results.
Output Format
What you receive when a job completes — depends on the model.
The output extension is announced by every model in its registry entry:
  • Segmentation & image-generation models return a .nii.gz volume in the same voxel space and orientation as the input.
  • Classification, captioning, and analysis models return a structured .json document with the model's predictions or numeric metrics.
NIfTI volumes are directly compatible with:
ITK-SNAP 3D Slicer FIJI / ImageJ nibabel (Python) SimpleITK
DICOM folders
DICOM folders are converted to NIfTI on upload; this section covers how that conversion behaves.
On the Submit Job page, choose select a DICOM folder. The platform inventories the folder into its series (multiple series per folder are normal — pick the one you need), converts the chosen series to .nii.gz, and adds it to your file queue — no re-upload.
  • Pseudonymization by construction. The converted NIfTI is built from scratch out of voxels, geometry, and spacing only — no DICOM header attribute (patient name, IDs, dates, device data) can survive into it. Uploaded DICOM files are deleted automatically after 30 minutes.
  • Geometry-safe. Slices are ordered by their true spatial positions; series with missing slices, mixed orientations, or non-uniform spacing are rejected with an explanation rather than converted into a distorted volume.
  • Current limits. Classic single-frame series, uncompressed (or RLE) transfer syntaxes, 2 GB per folder upload. Enhanced multi-frame DICOM and JPEG-compressed exports are detected and reported per series.
Nalvera Privacy Vault
Nalvera Privacy Vault is a free desktop app (macOS, Windows, and Linux; currently under development) that pseudonymises CT/MR scans fully offline before you upload them here.
It strips identifying DICOM header fields (patient name, ID, birth date, referring physician, institution) while keeping technical details like the scan date, with optional face removal, embedded-text removal, and end-to-end encryption — all switched off by default and run entirely on your own machine, with no automatic upload.
  • .nvlt will be an accepted input format. When encryption is turned on, the app exports a .nvlt container that only the Nalvera.AI platform can decrypt. Once Privacy Vault ships you'll upload it directly on Submit Job, in Study Design, or in the Interactive Studio — the platform already decrypts .nvlt automatically before the usual validation and inference pipeline runs, exactly as it would for a plain .nii.gz.
  • Encryption is optional. Without it, the app exports a regular .nii.gz scan file.
The Viewer
Tri-planar review of every completed segmentation, fully keyboard-operable. Research use only — not a medical device.
  • Navigation: click or Tab to a pane, then scroll, ↑/↓, PgUp/PgDn, Home/End; 1/2/3 select panes; press ? in the viewer for the full shortcut list.
  • Crosshair: Shift+click any pane to jump all three planes to that voxel, with its coordinates and true intensity (HU for CT) read from the raw data.
  • Window/Level: one-click CT presets (soft tissue, lung, bone, brain) or free sliders.
  • Label legend: searchable for large label sets, with per-label colours, show/hide, and bulk All/None that respects the active filter.
  • Editing & revert: the correction toolkit lets you fix a segmentation in the browser. The first save preserves the model's original output permanently — an Edited badge appears, downloads are named …_output_edited.nii.gz, and one click reverts to the model output at any time.
Provenance & Reproducibility
Every job records exactly what ran, on what input — citable in a methods section.
At submission the platform stamps the model version; at completion it records the Docker image, its content digest, and the worker that executed the job. Download the machine-readable manifest from the job's Provenance action on My Results, or via GET /api/jobs/<id>/provenance: 
{
  "schema": "nalvera-job-provenance/1",
  "job_id": "AB12CD34",
  "model_id": "totalsegmentator_ct",
  "model_version": "2.5.0",
  "docker_image": "totalsegmentator_v250",
  "docker_image_digest": "sha256:…",
  "executed_by": "gpu001",
  "edited": false
}
Results always display the version that actually ran, not the current catalog version. If a job fails, you can re-queue the retained input with one click (Retry job); credits for failed jobs are always refunded automatically.
Workflow Overview
What happens to a job after you click Submit.
Queued — validated, charged, and waiting; your live cluster position is shown on Job Queue. Processing — claimed by a GPU worker (cancellation is honoured at the next heartbeat). Completed — outputs and metrics appear on My Results instantly (pages update live). Failed — the reason is shown under the job row, credits are refunded, and Retry job re-queues without re-uploading. Each model page (e.g. TotalSegmentator CT) documents its training data, intended use, known limitations, and citations.
Return to Platform