Annotating

Annotators answer a project’s tasks in the workspace: the worklist on the left, the images in the middle, and the task panel on the right. The project’s manifest decides which tasks there are, at which level (patient, study, series or image) and with which labels. The workspace’s guide (the ? button, or the offer on a first visit) walks through the tasks of the item on screen, with their labels and keys. Series show a small thumbnail in the worklist when their pictures could be previewed at upload, and a task’s Guideline cards show the admins’ written guidance with example images of each option or label.

Annotation mode

The workspace opens in view mode: browse the worklist, use every viewer tool, and read the answers. To answer, press Annotate in the top bar (or Shift+A). Annotation mode hides the worklist and the side panels, shows only the open item’s tasks, and adds the tools of the tasks (brush modes, drawing tools) to the toolbar above the image; on a video the player pauses and stays on its frame. Next item stays in the top bar. Exit (or Esc, once nothing else is left to cancel) goes back to view mode: what is pending is saved at once, and in a project that submits on leaving your answers on screen are submitted. Annotation mode is part of the page’s address, so a reload or a Lock brings you back to it.

Work saves itself

There is no Save button. Every answer, shape and brush stroke is saved on its own, twice:

On this device, at once, encrypted under the project key: a reload, a closed tab, a browser crash, a lost connection, Lock, sign-out or an expired session loses nothing.
On the server, seconds later, encrypted the same way. Drafts the server has not received yet are sent when the connection or the session comes back, the next time the project is opened on this device.

The cloud in the top bar says where your work stands: saved on the server, being saved, kept on this device while you are offline, or a problem. Press it for the details: what is kept on this device only, each with the reason (the server refused it, its task left the manifest, it changed elsewhere meanwhile, or mask strokes wait for their series to open). Open takes you to the item, Retry sends a refused change again, and Copy and Discard are there for what cannot be saved any more.

Device drafts never leave the device unencrypted, and they open only with the project key. A draft written on another computer is on that computer until it reaches the server.

Submitting. There is no Submit button. By default each answer is submitted as soon as it is complete and valid, and editing a submitted answer revises it. Projects that run blinded multi-reader studies submit when you leave instead: answers stay drafts, saved and visible to you, until you leave annotation mode or the item, and you see the other readers’ answers only after submitting yours. Every change, before or after submitting, is kept as a new revision.

History. Each time you enter annotation mode and leave it is one session. In view mode, History at the bottom of the task panel lists the sessions that changed the open item: who, when, for how long and which tasks. Preview shows a session’s masks and shapes as they were at its end, and its other answers as text under it; Restore brings that state back as a new revision, including what a later session deleted; Delete this session reverts every change one of your sessions made, on every item, once it has ended, and brings back the shapes, masks and answers it deleted (delete later sessions first if you changed or deleted the same annotations again). A deleted session stays in the history, marked deleted.

Process log. While you annotate, your browser records how, for research on the annotation process: the items you open, your active and idle time, navigation, the tools you arm, mask strokes, which answers and shapes changed (never what they say), undo and redo, guidelines opened and AI tool use. It is encrypted like your annotations and stored per session. The project’s setup chooses everything (the default), summaries per item only, or nothing; admins can include it in exports.

Tasks and answers

Classifications and values: click an option or type the answer; number keys pick options of the focused task.
Required tasks must be answered before their section can be submitted (when you leave, in projects that submit on leaving) or before the item is done. Tasks may be shown only for some answers of another task.
Next item (the key the guide names) opens the next item that still needs your answers, required region answers included.

Shapes

Points, boxes and polygons are drawn on the images: pick the label, then draw. A box is dragged or placed with two clicks on opposite corners; a polygon is clicked corner by corner and closed on its first corner or with Enter.

Editing: select a shape, then drag a handle, or click a handle and then where it goes, or press the arrow keys (Shift moves further). The shape’s row in the task panel changes its label.
Holes: select a polygon, press Cut a hole and draw the hole inside it.
A shape in progress is saved like any other work: it is still there after a reload.
Undo and redo cover drawing, editing and deleting.
Whole-slide images (pathology slides: SVS, OME-TIFF, pyramidal TIFF, DICOM whole-slide images) open in the slide viewer: drag to pan, scroll or pinch to zoom, = and - to zoom and the arrow keys to pan from the keyboard, R to see the whole slide. Only the tiles on screen are read and decrypted. Shapes are drawn and kept at the slide’s full resolution, and measured in millimetres when the scanner recorded the slide’s scale. A DICOM slide is uploaded as its folder of instances: its levels become one series, and its label and overview pictures are left out, since they show the glass slide’s label. A stack of planes (confocal endomicroscopy, focal planes or channels in one TIFF) shows one plane at a time: Page Up and Page Down step between them, and a shape belongs to the plane it is drawn on. NDPI and MIRAX slides, and TIFF slides with JPEG 2000 tiles, must be converted to a pyramidal TIFF or OME-TIFF first.
Videos (MP4 and WebM) play in the video player’s bar: Space or K plays and pauses, ← and → (or ↑ and ↓) step one frame, Page Up and Page Down step between the frames a task annotates, J and L (and the ±1 s and ±10 s buttons) jump back and forward, and # goes to a typed frame number or time (125, 1:02.5). A phases task splits a video (or a signal recording) into consecutive phases: pick a phase and press Start here where it begins; it lasts until the next one starts. The timeline under the player shows a track per phase: drag a start, or focus it and press ← and →, to move it by one frame; right-click it (or Shift+F10) to jump there, change its phase or remove it; right-click a track to start that phase there; zoom with the buttons or Ctrl+wheel. Boxes and polygons on a video follow an object: draw it on one frame, then on a later frame copy it there (Copy the shapes of frame n here) and move it; on the frames between, the shape shows interpolated (dotted), and moving it there adds a keyframe. After its last keyframe the object stays where it was until you select it on the frame where it leaves the view and press Out of view from here. Copy the mask of frame n does the same for a mask, and Interpolate from frame n fills the frames between that frame’s mask and this one’s with the armed label’s shape blended from both (one undo step). A processor that propagates a mask through the video proposes one per frame; Start from this job’s masks on n frames starts your masks from them on every frame you have none of yet.
Ultrasound stills and cine loops (DICOM single and multi-frame images, or MP4 and WebM exports) open in the stack viewer; a cine loop plays in the video player’s bar at the speed the scanner recorded, with the same keys as videos. Shapes drawn inside a region the scanner calibrated (dashed outline) are measured in millimetres; elsewhere in pixels. Masks are painted frame by frame, as on videos.
Surface meshes (intraoral scans and other 3D surface scans: STL, PLY and OBJ) open in the mesh viewer: drag to turn the mesh, right-drag or Shift-drag to pan, scroll or pinch to zoom; the arrow keys turn it, = and - zoom, R shows it from the front. A point task places landmarks where you click the surface (drag a selected one to move it). A mesh region task paints regions of the surface with a brush: Shift erases, Ctrl starts a new region, [ and ] resize the brush. Coordinates are the file’s own (usually millimetres). Point clouds (meshes without faces) are not taken: export them again as a mesh first.
Signals (ECG and EEG recordings: DICOM waveforms, EDF, EDF+ and BDF files, PhysioNet WFDB records) open in the waveform viewer: one lane per channel over an ECG grid, with paper speed ([ and ]), gain (= and −) and display filters (high-pass, low-pass, notch) in the toolbar; drag or scroll to move along the recording. Arm a label with Mark with, then drag across the waveform to mark an interval or click to mark an event (Alt: on one channel only); drag an edge to move it, Delete removes the selected mark. Without a pointer, ← and → move the cursor, Shift+← → select a range, Enter marks it. Marks autosave like any answer.

Masks (labelmaps)

Segmentation tasks are painted with a brush:

Modes: Paint, Threshold (paint only within an intensity range), Erase, and Grow (fill the region within the range from the point you press). [ and ] change the brush diameter, in millimetres.
Volumes (CT, MR, NIfTI, NRRD): paint on any plane. Interpolate slices fills the slices between two slices you painted.
Single images (photographs, radiographs, one DICOM image): a task at the image level gives each image its own mask, painted on the stack viewer. On an image with no pixel spacing the brush diameter is in pixels. Interpolation does not apply.
Display: “Show my mask” and the mask opacity are in the task panel.
Touch and stylus: one finger or a pen draws and paints; two fingers zoom and pan. A pen’s pressure sets the brush diameter.

Region tasks

A region task asks about each shape or each labelled mask of another task, for example the type of every lesion you outlined. Select a shape (or a saved mask): its region tasks appear under it in the task panel. A dot marks a shape that still misses a required answer, and submitting asks for those answers first. Deleting a shape deletes its region answers.

Processor proposals

A processor (a model the project runs) may propose answers, shapes and masks. They appear under Processor proposals in the task. Each proposal is one card: Use this answer takes it over as your own answer, which you correct as you like, or keep unchanged to confirm it (for a mask, Show draws it and Start from this copies it into your mask). Nobody accepts or rejects proposals: a member’s answer made from one is what counts. Your browser checks each proposal’s signature against the processor’s registration before showing it: one that does not verify says so and cannot be taken over.

Dictation

When the project binds a transcription processor to a free-text task, the task shows Dictate. Turn it on, then hold Hold to talk (or Space on it) while you speak: the text appears as you talk, and when you release you get the text as a proposal. Keep adds it to your answer, which saves like typed text and can be edited; Discard drops it. The browser asks once for the microphone, which is open only while you hold the button.

The processor receives your voice and none of the project’s data, and declares that it keeps no audio. The panel says before you record whether the task keeps your voice: by default only the text is kept; a task set to keep recordings stores each kept dictation’s recording, encrypted, with your answer, where you can play it back, and exports it.

Correction and opinions

A task can add a correction stage. Its readers first answer each item on their own; once every reader of an item has submitted, the task’s other annotators find the item in Next item, and the task lists Readers’ annotations to correct. Start from one with Use this answer (Start from this for a mask), correct it, or keep it unchanged to confirm it: your annotation is the correction, and the latest correction is the task’s value on the item. Readers do not correct their own items. When the task splits its items, it can also assign correctors per item: then only an item’s assigned correctors find it and correct it.

Opinions in the task panel shows every reader’s submitted answers, shapes, masks and their region answers, the corrections and, for tasks with several readers, the consensus, which project admins record. Nothing there is accepted or rejected.

Reader studies

A task can carry research options, chosen when the project is set up:

Confidence: a scale next to the task (for shapes, under each shape) asks how sure you are of your answer, from 1 to 5.
Known answers: what a project admin answers on such a task is a known answer, which annotators never see and against which the readers are measured. The task panel tells admins when their answers are known answers.
Randomised AI help: smart segment and processors’ proposals are offered on some items only, chosen at random but fixed for every reader, so that readings with and without them can be compared. On the other items you see processors’ answers once you have submitted your own.
Hidden repeat items: some items, chosen at random but the same for every reader, come back later in your queue for a second, independent reading, without being marked as such, so that each reader’s answers can be compared with their own earlier ones. Only the task read again is shown, empty (no shapes or mask of yours either), with its confidence question if it has one; your second answer is kept beside the first, and only you and the project’s admins see it.