Scientific Image / Volume

Scientific Image presents native numeric planes, stack and channel navigation, aligned segmentation labels and retained rectangle, polygon, freehand and volume-box regions of interest. The model and native runtime own every scientific result. The viewer changes display brightness and navigation only.

The image opens as a whole-plane overview. Nearby slice and channel sliders include previous/next buttons. Zoom in, Zoom out and Fit image change the view only; Show labels toggles segmentation outlines while leaving their native values available for inspection. Source setup, identities, units and advanced axes are tucked into Source and view settings. An unbound surface offers Choose source directly. Setup offers available arrays and tensors from the current model, while retaining a symbol field for sources not listed.

Bind a native rectangular array or a Tensor symbol in source.symbol. For an existing IMAGE_* result, author IMAGE_TO_ARRAY(image) in the model and bind that numeric plane. The surface reads resident model values or explicitly selected, byte-pinned PNG/TIFF/NIfTI Files sources. Sample decoding stays native.

The existing native IMAGE_DECODE accepts PNG bytes, including a BYTES_FROM_BASE64 value. Bind IMAGE_TO_ARRAY(IMAGE_DECODE(bytes)) to inspect that result. This decoder converts to normalized RGB8 before grayscale projection: it does not preserve scientific bit depth, alpha, acquisition metadata or physical calibration. Its model operation decodes the complete image; the viewer's bounded inspection does not bound that decoding cost. The canonical model includes a tiny teaching PNG for this normalized formula path. The separate original-file importer preserves raw PNG/TIFF samples and header-scaled scalar NIfTI volumes. DICOM Part 10 monochrome frames are imported separately from spatial volumes.

Review!type = "scientific-image"
Review!config = <toml>
  version = 1
  kind = "scientific-image"
  title = "Microscopy review"
  [source]
  symbol = "Volume"
  mask = "Labels"
  units = "normalized intensity"
  [axes]
  row = 2
  column = 3
  stack = 0
  channel = 1
  [calibration]
  rowSpacing = 0.5
  columnSpacing = 0.5
  unit = "µm"
</toml>
Review!data = <json>
  {
    "version": 1,
    "regions": []
  }
</json>

Axes are zero based and distinct. Omitted row/column axes use the native inspection default, the final two dimensions. Stack/channel axes name fixed dimensions. Other fixed dimensions remain individually navigable. Configure different row/column axes to inspect another orthogonal volume plane; the current slice is never a reconstructed 3D image.

Inspect and measure

Import original PNG/TIFF/NIfTI samples

1. Open Source and view settings and choose Original image file. Original bytes are retained in this saved model's Files under a unique path. Native inspection verifies the checksum and prepares the configuration.
2. For a TIFF, open TIFF page, enter its zero-based page and choose Inspect selected page. Only that page's pixels are decoded; failed inspection retains the uploaded original for another page selection. For a volume, select Import consecutive pages as a stack, enter Page count and choose Inspect selected stack. The prepared axes keep pages and channels separate; TIFF pages must have matching sample semantics.
3. Review dimensions and sample type, then Save image configuration. The acknowledged save applies path, SHA-256, starting page and optional page count. Failed/conflicting saves retain the draft. Supply calibration and units in advanced settings as needed.
4. Browse channels and inspect/measure rectangles or outlines as usual. Raw sample encoding is retained, including explicit alpha/extra samples. RGB channels remain separate planes; associated alpha is not unpremultiplied. TIFF orientation is recorded but not applied. Original sample identity exposes the encoding, stored orientation and extra-sample tags.

The canonical OriginalFile workspace starts unbound. Try the included 2 × 2 PNG16 fixture: its declared row-major values are 1, 257, 4096 and 65535, independently of the normalized PNG formula example. For stack navigation, try the included 2 × 2 × 2 TIFF16 fixture with start page 0 and count 2; its declared planes are [1,257,4096,65535] and [10,20,30,40].

Import a NIfTI volume

Choose an original .nii or .nii.gz through Original image file. The native importer reads the complete single-file NIfTI-1 scalar volume; there is no page-count setup. Save the prepared configuration, then browse its slices and draw regions or volume boxes using the same controls as a TIFF stack. The included 2 × 2 × 3 NIfTI fixture has stored values 0–11, header-scaled intensities −10–12 and spacing 0.5 × 0.75 × 2 mm.

Native import supports little- and big-endian signed/unsigned 8/16/32-bit integers and 32/64-bit floats. Header slope/intercept scaling is applied once to the selected decoded samples; a zero slope disables scaling and ignores the intercept. Original sample identity retains the original encoding, scale, voxel spacing, units, intent and both spatial transforms. Source checksums bind the original compressed or uncompressed file bytes.

Columns follow stored X, rows Y and slices Z. Import does not reorient, resample or assign anatomical labels. An orthogonal S-form supplies physical axis spacing; otherwise the native Q-form/pixdim spacing is used. Known units and orthogonal geometry prepare row/column calibration for review before save. Use NIfTI depth spacing, inside physical volume calibration, copies the native depth spacing only when saved row/column calibration still agrees. Sheared or degenerate S-forms retain their metadata without automatic spacing calibration; manually supplied spacing remains your responsibility.

Admission is 16 MiB for both encoded and decompressed bytes, at most 1,000,000 voxels for the complete volume and 1–8192 along each spatial axis. Higher reported dimensions must be singleton: time series and vector volumes require separate import. NIfTI-2, paired header/image files, complex/RGB samples, 64-bit integers and unsupported spatial units refuse. Gzip framing/checksum and header extension lengths are checked; extension contents stay uninterpreted. Inspection explicitly decodes the bounded selected volume each time; no background indexing, worker or cache is introduced.

Precision-preserving Files source admission

Selected PNG8/16 and TIFF pages may become byte-pinned Files sources. Import admits at most 16 MiB encoded, 8,192 per dimension and 1,000,000 decoded samples, with at most four channels. PNG palette, low-bit-depth expansion, transparency expansion and animation refuse; TIFF admits raw BlackIsZero/RGB samples in stored row order, without orientation or color transforms. Selected TIFF pages are zero based, limited to 0–63; preceding directories are traversed without decoding their pixels. Integer samples through 32 bits and IEEE floating samples retain their precision; 64-bit integer samples refuse rather than lose bits.

Import hashes the selected File once and uploads its original bytes through Files. Every native inspection and measurement re-reads and hashes the bounded asset before decoding the selected page; a checksum mismatch refuses. The surface configuration pins path, checksum and page. Physical calibration is never inferred. A single page has channels, not spatial depth and refuses volume boxes. Advanced TIFF page-stack import admits the separate depth workflow below. Segmentation comparisons remain unavailable for Files sources. Optional costs are admitted only by the selected Files-source request: up to 16 MiB encoded input plus bounded decoder buffers and two 1,000,000-sample numeric buffers. Ordinary model/Matrix inspection and the normalized IMAGE_DECODE/CV lane perform no scientific-file work. The core library keeps scientific-codec dependencies behind a default-off feature; the runtime-host product selects them at build packaging. No timing/RSS ratchet or full host compilation is claimed by these functional checks.

TIFF stack admission

Advanced import may select 2–64 consecutive TIFF pages whose start plus count fits within the first 64 directories. A stack retains separate page and channel axes. All selected directories must agree on dimensions, channel count, sample encoding, photometric interpretation, orientation and alpha/extra-sample tags; metadata is admitted before pixel decoding. Total samples across all selected pages remain at most 1,000,000 (8 MB native numeric storage), rather than a per-page allowance. Encoded bytes stay capped at 16 MiB. Decoder intermediate buffers are capped at 16 MiB and decoded buffers at 8 MB; at most one page is being decoded alongside the admitted stack. Projection admits a second bounded numeric buffer. No unselected pixels or ordinary Matrix/evaluator work is added. The source identity retains all-byte checksum, starting page and page count. Missing pages or any inconsistent page refuse the complete request. Each native inspection rehashes the complete original asset and decodes the selected bounded stack; there is no background decode or cache worker.

Page order is the user-selected depth order, not inferred physical anatomy or spacing. Native volume boxes admit at most 1,000,000 voxels over the page axis with other axes, including channel, fixed. Physical volume needs user-supplied page spacing in addition to row/column spacing. First slice is relative to the selected stack; the browsing controls also display the original TIFF page number. PNG and single TIFF page sources remain channel planes and refuse volume boxes.

Volume box admission

A volume box explicitly selects one fixed depth axis, a starting slice and positive slice count. Row, column and depth extents must fit within the native shape; the complete box contains at most 1,000,000 voxels. Choose Use complete box to select the full extent on these three axes, or adjust the folded exact coordinates. Other dimensions stay fixed. native admits the resident layout and checked box bounds before the first sample, then streams every selected voxel through the existing native counts, mean and finite-sample sum accumulator. The accumulator retains constant-sized statistics state, not a copied volume or JSON cell buffer. The returned display window remains at most 64 × 64; its size never substitutes for the exact measured box. Saved records retain the full box bounds and native counts, with the existing source, revision, calibration and original-file identity checks. Optional physical volume requires authored row/column spacing plus positive depth spacing in the same linear unit. Volume is the complete voxel count times all three spacings, including voxels with unavailable intensities. Other axes, including channels, stay fixed.

With complete positive calibration, Intensity-volume integral is the native raw intensity sum multiplied by row × column × depth voxel spacing. It is separate from integrated intensity, the finite-sample raw sum, and geometric volume, which includes every voxel. An integral is available only when every selected voxel is finite numeric. Blank, error, nonnumeric or nonfinite samples make it unavailable; they do not silently contribute zero. Missing calibration, overflow or underflow to a zero nonzero result also makes it unavailable. A complete all-zero or signed-cancelling volume may correctly have a zero integral.

The result retains its native value, per-voxel physical volume and spatial unit (e.g. mm³). Display combines that spatial unit with the existing authored intensity-unit label, or “raw intensity” when no label is supplied; the label is not dimensional certification. Save, reopen, print and retained-draft JSON export preserve the exact result or its explicit unavailability alongside the raw sum and geometric volume. Older records without this field remain readable.

This cost is selected only by the volume inspection request. Ordinary numeric inspection, plane measurements and the evaluator perform no volume traversal. Volume requests preserve the viewed working-revision fence and acknowledged record writes. This is a bounded box measurement, not reconstruction or a frozen snapshot of volatile/external values.

Polygon and freehand region admission

Polygon and freehand outlines are closed paths with 3–64 vertices inside one 64 × 64 native pixel window. Their native measurements include pixel centres inside the outline by the even-odd rule, including centres on the boundary. Self-crossing outlines use the same rule. Area is the number of selected pixels times authored pixel spacing, not a continuous geometric polygon area. Empty pixel selections refuse measurement. Outlines and native results retain the same source, revision, calibration, label and acknowledged-save ownership as rectangles. Outlines are bounded input geometry; scientific membership, counts, mean, sum and calibrated area are computed by native.

The optional outline selects its additional work only at image-inspection ingress: at most 64 vertices × 4,096 pixel-centre checks. Ordinary numeric projection and the ordinary evaluator perform no outline work. The viewer keeps at most 64 drawing points, with explicit refusal when that bound is reached; the existing 200-record/300 KB document and 4 KB request bounds still apply.

1. Select a sample to read its exact native value and optional segmentation label. The overview contains at most 48 × 48 native samples at regular row and column strides; it performs no averaging and may omit small features. Select Inspect selected area to navigate directly to a 24 × 24 detail window containing every native pixel. Detail paging appears for larger planes; Whole image returns to the overview. Arrow keys move between samples with a single tab stop. Folded Go to pixel accepts an exact zero-based row and column within the current plane, requests native detail, and selects that pixel without paging through the intervening image. Slice/channel controls request another plane. When only the slice, frame or channel changes within the same native grid, the selected stored row/column stays selected and shows the fresh native value. Changed source, shape, view axes, window or sample stride clears selection. This follows a stored coordinate, without image registration or a claim that the same feature persists between frames.
2. Open Measure a region and choose Select rectangle on image. Select one corner and then its opposite corner; the outlined rectangle supplies the coordinates automatically. A single selected pixel can also anchor a measurement. Keyboard users move with arrows and press Enter to choose each corner. Exact coordinates and size exposes optional numeric edits and permits regions up to 64 × 64 pixels. Select Measure native region. The native endpoint returns finite, zero/nonzero, blank/error/nonnumeric/nonfinite counts and minimum/maximum for exactly that rectangle, up to 64 × 64 pixels. It also returns finite-sample mean, integrated intensity (the sum of finite samples) and rectangle pixel area. Samples marked blank, error, nonnumeric or nonfinite are excluded from mean and sum, with their counts retained. An unavailable or overflowing sum is shown as unavailable, never fabricated. Optional Pixel calibration supplies known row/column spacing in m, cm, mm, µm/um or nm. Native rectangle area multiplies the full rectangular extent by that spacing, including pixels with unavailable intensities. Spacing is authored metadata for the configured display axes; it is not inferred from an image or certified by the viewer. For an irregular region, change Region shape to Polygon or Freehand and select Draw. Select polygon vertices in order, then Finish outline, or drag a freehand boundary and release to close it. Keyboard users move between native samples with arrows and use Enter to add vertices, then select Finish outline. Undo point and Continue outline keep the geometry editable before measurement. Drawing snaps to 1/16-pixel coordinates, retains at most 64 vertices and explicitly stops accepting more points at that bound. Review the visible outline before finishing. Its enclosing native window must fit within 64 × 64 pixels. Native counts, mean, integrated intensity and calibrated selected-pixel area include only pixel centres inside or on the closed outline. The underlying returned cell window still covers the enclosing rectangle. Self-crossing paths use even-odd inclusion. A path selecting no pixel centres refuses measurement. Selecting details first makes small boundaries easier to draw; the overview may omit small features. For a region spanning multiple slices, choose Volume box, select a rectangle, and supply First slice and Slice count. The starting slice is zero based. Volume depth axis selects a fixed spatial dimension; configured channels stay fixed. Choose Use complete box for the full extent on the selected axes. The native box admits at most 1,000,000 voxels in total; exact coordinates remain available for smaller boxes. native returns counts, minimum, maximum, mean and finite-sample sum for the complete box, without averaging or aggregating planes in the browser. Physical volume calibration accepts known depth spacing in the same unit as the configured row/column spacing. Leave it empty for voxel counts only. Calibrated volume includes every voxel in the box, including unavailable intensities; it is geometric extent volume, not an intensity integral. The visible native plane is a preview of one slice. Whole-box statistics are recorded separately from that plane's counts.
3. Enter a label and select Save measured region. The record preserves the source symbol, source hash, native dimensions, axes, fixed indices, rectangle, outline or volume box, authored intensity units, model working revision, native measurements and any authored calibration. Up to 200 records are retained, with 20 displayed per page. Open Saved regions to revisit them. Search saved regions searches the complete retained history by label and recorded source or mask identity before paging. Clear the search to browse all records; print and complete exports always include every record. Show region returns to its slice and drawn boundary when the authored source, dimensions and displayed axes still match, including in Run mode. The whole-plane outline shows the retained geometry. Detailed source identity and counts stay available in each record's disclosure. In authoring mode, Measure again copies the saved rectangle, outline, selected mask label or volume box into an editable new draft. The original record stays unchanged. Adjust the size, slice interval or outline, review current calibration, then measure natively before saving a separate record. Historical values are never reused as a new measurement. The current source hash, shape, mask and admitted plane must match; changed input revisions are measured freshly. Drafts retain their source/shape/plane ownership across browsing and refuse measurement after owner or native shape changes. An edited volume interval can move beyond its historical slice range; other fixed axes stay attached to their recorded indices. Returning to an outline restores both recorded display axes along with its indices and origin. Return to saved region plane restores the draft's recorded axes and indices. Volume depth spacing is copied only when its saved unit matches the current reviewed plane unit; otherwise supply reviewed spacing again or measure voxel counts only. Volume records also retain their depth axis, starting slice, slice count, voxel count and any physical-volume calibration. Their rectangle highlights on every slice within the recorded interval; other axes remain fixed.

The optional source.mask uses the existing native comparison operation, which admits matching dimensions and returns both values in one response. Labels can have different semantic types or units from the intensity source. Every label must be a finite nonnegative integer. Nonzero labels are outlined; the selected pixel shows its numeric label. Missing, malformed or misaligned masks refuse inspection instead of displaying an unverified overlay. This surface inspects segmentation produced by native model formulas; it does not run segmentation in the browser.

Brightness initially uses the overview's native sampled min/max and stays fixed while navigating slices and detail. Contrast provides editable display limits and Use current view range when a new range is useful. These presentation controls do not change data or compute a scientific result. Intensity units are authored metadata; coordinates remain pixels. Arbitrary 3D outlines, unbounded paths, compressed/enhanced DICOM and NIfTI-2 codecs, raycasting, surface reconstruction and learned automatic segmentation are explicitly unsupported. Integrated intensity is a finite-sample sum. A complete, calibrated volume box separately records its native intensity-volume integral; plane sums do not imply a physical integral.

Export saved measurements

Download measurements exports CSV for every saved ROI and every label in saved whole-mask or spatial-volume captures, including records beyond the current page. It works in authoring and Run even when the live source is unavailable. Values are exact stored native results: pixel/voxel counts, mean, raw intensity sum, physical area, geometric volume and calibrated intensity-volume integral use distinct columns. Empty numeric cells stay empty for unavailable or unrecorded values; sum/integral state columns distinguish those cases. Zero remains zero. No values are remeasured, averaged or calculated in the browser.

CSV retains each record's source/hash/revision, mask/label and intensity units. Its geometry column preserves native bounds, axes, indices, calibration and original-file metadata where recorded. Capture scope distinguishes image labels from spatial-volume labels. Authored text is quoted and protected from spreadsheet formula interpretation; numeric values retain their stored precision.

Open Export options → Download exact records (JSON) for the complete admitted image document, including full source provenance and every native capture. These are historical records, without a claim of current live freshness. Export excludes unsaved measurement/capture drafts, preserves originals and needs no write access. Uncertain-write recovery keeps its separate draft download.

Save ownership and recovery

Retained region writes use the host's data slot compare-and-swap revision and must return a matching ready revision containing the requested document. Failure or conflict keeps the label, outline and measured draft. Browsing and Refresh also preserve drawing and measurement drafts. Return to outline returns to a drawing's original plane after browsing another slice; changed source ownership requires discarding that geometry before drawing again. The draft identifies when it belongs to another slice. Inspection navigation is disabled during measurement or save, and unfinished drawings protect navigation away from the surface. Volume-box parameters and measured results survive browsing and failed saves. Changing source ownership retains the volume draft and requires explicit discard before selecting a box in the new source. Removing another saved record is disabled while a drawing or measurement draft is open. Changed authored source retains the draft but requires remeasurement before saving. An ambiguous or failed write locks another save until explicitly resolved. Keep draft and stop waiting releases the waiting screen without issuing another write. Observe saved regions reads the saved data: an advancing, matching copy confirms the original save. A different saved copy keeps the draft available for download and offers Use saved regions and keep draft; remeasure before adding the retained measurement to that copy. Late responses after stopping cannot clear the draft. Configuration edits similarly require the host's acknowledged source-write capability. Read-only hosts show inspection without offering mutation, and hosts lacking configuration acknowledgements can use their existing source editor.

A changed authored source marks retained measurements as historical. For model sources, the source hash identifies authored source, not each live input value. Imported-file source hashes identify original bytes and selected page; native reads recheck bytes. Replacing file bytes requires explicit reimport, and saved regions retain the original byte identity. Even records with the same source hash need remeasurement after input changes. Native inspection captures the model working revision before and after reading. Measurement requires the viewed working revision and refuses recorded input/source changes. That token does not freeze volatile formulas or independently changing external values; model-source hashes remain authored-source identities. Refresh the plane and measure again to add a new reviewed record; historical results are never silently recomputed. Removing a region also requires the guarded receipt. Print includes a bounded native sampled plane, provenance and every admitted retained region (up to 200), without writes or navigation controls. Older native hosts lacking the selected image sampling/measurement contract refuse inspection instead of silently returning weaker results.

See the canonical model and implementation/evidence scope.

Imported TIFF stacks and NIfTI volumes offer XY, XZ and YZ stored planes. The slice control follows the remaining spatial axis while channels stay separate. Selecting a pixel before switching planes carries its stored coordinates into the next plane. These views neither resample voxels nor infer anatomical directions. Saved regions retain their row/column axes and reopen on their original plane. Discard an unfinished rectangle before changing its plane or slice; measured records keep their recorded provenance. Confirmed NIfTI orthogonal calibration follows the selected axes. TIFF cross-plane measurements stay uncalibrated unless spacing is explicitly supplied for those axes; row/column spacing is never guessed as stack spacing.

Import stored DICOM frames

Select an original .dcm or .dicom Part 10 file. The native importer admits uncompressed implicit/explicit little-endian and explicit big-endian MONOCHROME1/2 samples with 8- or 16-bit allocation and signed stored bits. Header slope/intercept produce the inspected intensity values. MONOCHROME1 inverts presentation brightness only. Pixel spacing, when present, prepares row/column calibration in millimetres for review before saving the configuration.

Try the synthetic 2 × 2 × 2 frame fixture. Its header rescales intensities by 0.5 with offset −10 and supplies row/column spacing of 0.5 × 0.75 mm. Use Frame to browse the stored frames. Header position, orientation, slice thickness and inter-slice spacing remain metadata; they do not prove a spatial stack. These frames have no XY/XZ/YZ selector or volume-box measurements. Padding values remain included in native statistics and region measurements, with padding metadata disclosed alongside provenance.

Only selected technical image metadata is surfaced; patient/demographic tags are not projected into the viewer. Original files remain in Files unchanged. Compressed transfer syntaxes, enhanced functional groups, LUT/reference transformations and floating Pixel Data refuse. No anatomical reorientation, reconstruction or clinical interpretation is performed.

Reviewed plane spacing also sets display proportions, including cross-plane NIfTI views. Fit and zoom preserve those proportions while native pixel selection, drawing and ROI coordinates remain in stored pixel space. Without reviewed spacing, the display uses square-pixel proportions. Display scaling does not resample values or change native measurements.

Create a model-backed threshold mask

For a named two-dimensional model image, open Segmentation and choose Create threshold mask. Enter a threshold in the original intensity units. Pixels equal to or above it receive label 1; blank, error and nonnumeric samples receive zero. The formula is saved under a fresh model name and follows source values through the existing native evaluator. Source intensities and existing masks remain unchanged.

Mask options offers native opening to remove small bright features and closing to fill small dark gaps, with radius 1–5 pixels. These operations act on binary labels, without intensity normalization or physical resampling. The ready native image shape supplies the formula dimensions; guided authoring admits at most 1,000,000 pixels on stored row/column axes. Files and volume masks continue to require separately authored native model data.

After the native formula receipt and aligned mask inspection, the viewer prepares its existing image configuration. Save image configuration connects the overlay through the acknowledged configuration owner. A failed inspection keeps the created mask and offers Observe saved mask without another source write. An ambiguous write cannot be repeated: keep the threshold draft, observe the exact saved source, or review it in Code. Keep draft and stop waiting retains parameters and ignores late completion; it does not undo accepted work.

Measure a segmentation label

With a model-backed mask connected, choose a rectangle or finish an outline under Measure a region. Open Measure labeled pixels, enable Restrict the area to one mask label, and enter the label (1 is the usual foreground of a threshold mask; 0 is background). Use selected pixel label copies the label already inspected at the chosen pixel. Measure native region counts only that label inside the exact bounded area. The native result supplies selected-pixel count, physical area from reviewed spacing, intensity mean, integrated intensity, extrema and missing-value counts. Other labels contribute nothing to those measurements.

A label absent from the area reports zero pixels and area, with unavailable intensity summaries. Area counts every selected pixel, including pixels whose intensity is missing; means and integrated intensity include finite intensities only. The operation admits at most a 64 × 64 window, with outlines retaining their existing 3–64 vertices. These are regional measurements, not an estimate of the whole mask from a sampled overview.

Label the result and save it through the existing guarded region writer. The record retains its mask symbol, integer label, working revision, source identity, axes and exact bounds. Show reopens the recorded region and focuses that label in the live overlay when the same source and mask remain connected. A record from a different mask stays historical. Print includes mask provenance and every retained result. Imported-file comparisons and labeled volume boxes remain unsupported; stored DICOM frames are never treated as a spatial volume.

Analyze the complete mask

For a model-backed 2D image with an aligned label symbol, open Whole-mask analysis and choose Analyze whole mask. The native result counts every pixel for each label, including background zero, and supplies reviewed physical area, finite intensity mean and integrated intensity, extrema and missing-sample counts. The table shows 20 labels per page. Download summary retains the complete native capture, its model, source hash, image and mask symbols, actual shape and working revision, plus authored intensity units.

This explicitly selected operation scans the complete image rather than the sampled preview or an ROI. It admits at most 1,000,000 pixels and 256 distinct nonnegative safe-integer labels, including background. Existing native dense array admission avoids per-pixel structural serialization; generic arrays retain their 100,000 structural-cell admission budget. Invalid mask samples, excess labels, changed revisions or mismatched shapes refuse the entire result. Volumes and imported-file comparisons remain outside this operation.

Label counts and physical areas include pixels whose intensities are blank, errors, nonnumeric or nonfinite. Intensity summaries include finite samples only; missing summaries and overflowed sums remain unavailable. Reviewed row and column spacing applies to stored axes 0 and 1. Labels group all pixels with the same integer value; they do not split disconnected components. The existing native IMAGE_COMPONENTS formula remains available for image connected-component geometry. Captures describe the recorded revision: analyze again after changing live inputs. Whole-mask captures can export as JSON or save into the image document; saved document regions remain separate exact bounded ROI measurements.

Save and compare whole-mask captures

After analysis, review the native totals, edit Capture name if desired and choose Save whole-mask capture. The image document retains up to five complete captures within its shared 300 KB budget, alongside existing regions. Each capture includes the exact native summary, model/partition, source hash, image and mask symbols, shape, working revision, authored intensity units and reviewed calibration. The default name includes the recorded revision.

Open Saved whole-mask captures, then Open a capture to inspect its recorded label table. Reopening reads the saved document without recomputing or connecting a historical mask. A current-match message requires the viewed native source hash and working revision, mask, shape, model/partition, units and reviewed calibration to agree. Otherwise the capture is explicitly historical. Print includes every saved capture and all its labels.

When compatible captures exist, Compare recorded captures pairs the saved before/after pixel counts and intensity means. Areas and integrated intensities sit under an advanced disclosure. Only captures with the same model/partition, source hash, image/mask symbols, shape, units and calibration are comparable; their recorded working revisions may differ. All scientific values remain the native recorded values; no browser-derived differences or reanalysis occur.

A save requires a matching ready receipt at an advancing document revision. An uncertain write retains its draft, supports the existing Observe saved regions recovery and full-document download, and cannot repeat automatically. Changes to the native source or document revision preserve the local capture for download and require fresh analysis before saving it. Region and capture removals preserve all other records. Read-only Run supports analysis, reopening and download; capture writes require the authoring host.

Analyze a declared spatial volume

For a native model image and aligned integer mask, open source settings, then Labels, units and axes → Spatial volume declaration. Explicitly review the three axes as orthogonal spatial depth, row and column. They must cover the three-dimensional native shape and match the configured stack/row/column axes; a channel axis or original-file binding refuses the declaration. Tensor rank does not identify spatial meaning. Stored frames, channels and time sequences must not be declared spatial merely because they have three dimensions.

The active declaration remains visible beside the image. Open Whole-volume label analysis and choose Analyze volume to scan the complete admitted volume: up to 1,000,000 voxels and 256 integer labels including background. Native voxel counts, finite intensity summaries and missing-sample counts cover the entire declared volume, independently of the displayed plane or ROI box.

Review positive row/column pixel spacing and optional depth spacing in the same linear unit for physical volume. Without depth spacing, the native request omits calibration and every physical volume remains unavailable, even if the displayed plane has reviewed pixel spacing. Saved captures retain the native null calibration and declared axes, alongside any authored plane-spacing metadata. No partial spacing is used to manufacture a physical volume.

Name and Save volume capture to retain the exact native summary with reviewed geometry, full native calibration or null, source/mask/shape/revision provenance, units and model/partition. Image and volume captures coexist under the same five-capture/300 KB document limit and guarded write/recovery flow. Reopen saved captures without reanalysis, print every label, or compare paired native values when scope, axes, depth spacing, plane calibration and remaining source identity agree. A native working revision match is additionally required for the viewed current-match message. Pending analysis is discarded when its source, geometry, calibration, partition or capture-document revision changes.

See Native spatial-volume label analysis and the canonical spatial example. The declaration records the author's reviewed geometry; it does not authenticate specimen or medical geometry, reconstruct a DICOM stack, apply an affine, resample voxels or reinterpret channels as depth.

Label disconnected foreground regions

For a model-backed 2D image with a connected binary mask, open Segmentation → Label connected regions. Create component labels authors a fresh native IMAGE_LABELS formula and shows its bounded native label preview. The source image and original mask remain in the model. Native zero remains background; positive component IDs follow the first foreground pixel in stored row order. The default connects pixels touching an edge. Mask options → Connectivity can also connect pixels touching a corner, and the fresh mask name can be edited there. The formula stays reactive: IDs may change when foreground pixels change.

Review the preview, then choose Use component labels. A fresh native revision check precedes the prepared image configuration. Save image configuration is the explicit step that connects the component-label overlay. Analyze the whole mask afterward to obtain exact native per-component counts, physical areas and intensity summaries, or retain a labeled ROI/whole-mask capture.

The input must contain only finite numeric 0/1 samples. Existing label IDs greater than one, missing samples or other values do not silently become foreground. Create and connect a threshold mask first; the guided threshold workflow uses original intensity units and can apply optional native opening/closing. Labeling admits up to 1,000,000 pixels and 255 foreground components, plus background, so complete summaries remain inside the existing 256-label admission. Native validation checks the full mask and refuses invalid or excessive components without a partial result. Volumes, Files, channels and inferred spatial axes do not enter this two-dimensional action.

The source writer checks a fresh viewed native source hash and working revision before creating a formula, requires an acknowledged ready source receipt, and checks that receipt's native source identity when inspecting the new mask. Connection rechecks the preview's native working revision. A changed source or failed inspection retains the created formula for review; observe the saved mask and inspect again rather than repeating its write. An uncertain write cannot automatically repeat. The preview samples pixels and supplies no browser-derived component counts, geometry or intensity statistics.

Whole-mask and volume analysis temporarily hold region editing and source changes. Save or discard the capture draft before starting a region, so saving one kind of measurement cannot discard another. Late native results are ignored when the source, calibration, partition or saved-document revision changes. Historical captures and regions can still be removed through the guarded document save when their live source is unavailable.

Use imported samples for segmentation

Open Segmentation → Use plane as model data on a pinned PNG, TIFF, NIfTI or DICOM plane. Create model image re-reads the original file in native and copies every pixel on the selected stored plane into a fresh 2D model formula. The selected channel and frame stay fixed. This is an immutable copy: changing Files, frames or live inputs does not update the imported samples. The original file remains in Files.

The complete selected plane admits at most 1,000,000 finite numeric pixels when the exact authored model source fits 2 MiB in UTF-8 bytes. Larger or nonfinite planes refuse without sampling, truncation or replacement. native authors the literal and a Code comment retaining original SHA-256, byte count, technical metadata, source shape, stored axes/indices, reviewed calibration and preparation revision. The browser transports that formula through the existing acknowledged source writer; it neither decodes samples nor computes scientific results.

Review the native saved preview, choose Use model image, then Save image configuration to connect the copy. Reviewed plane spacing is retained on the new row/column axes. Create a threshold mask and component labels using the existing segmentation controls; native whole-mask counts, areas and intensity summaries and saved captures then work on that durable model image. No spatial volume is inferred from an imported file stack or DICOM frames. An uncertain write retains its draft for Observe saved model image and is never repeated automatically.

Label disconnected regions in a declared volume

A model-backed rank-3 image with reviewed spatial axes and a connected binary mask offers Volume segmentation → Label connected volume regions. Create volume labels authors a fresh native mask, previews the selected stored plane, then Use volume labels prepares an explicit image configuration save. The original source and mask remain in the model.

IMAGE_LABELS_3D(binary, size0, size1, size2, depth_axis = 0, row_axis = 1, column_axis = 2, connectivity = 6) preserves the stored shape. Reviewed spatial axes must be a permutation of the three stored axes. Foreground IDs follow the first depth/row/column encounter; zero stays background. The default uses touching faces. Folded connectivity options admit faces and edges (18) or faces, edges and corners (26). These are voxel adjacency rules; spacing does not resample or certify geometry.

Native labeling checks the complete flat or rectangular nested rank-3 binary mask against the stored extents. It admits at most 1,000,000 voxels and 255 foreground regions plus background, and refuses missing/error/string/nonfinite samples or existing labels above one. Create a strict native volume threshold mask before labeling when the source contains intensities rather than binary labels. Analyze and retain the resulting labels using the existing exact whole-volume capture workflow.

Try the connected spatial volume example to compare face and corner connectivity with a reviewed permutation of stored axes.

Copy a complete file stack into model data

Single-channel retained rank-3 TIFF stacks, NIfTI volumes and DICOM frame collections offer Copy full stack as model data. Create model stack copies every admitted stored sample, preserving the complete stored shape and ordering. Copy preparation checks the exact retained starting page/count and any returned output axes/indices for both scalar and channel copies. Legacy preparation receipts without optional output metadata retain their original axis defaults. The selected plane and fixed indices determine the saved native preview; they do not select a subset of voxels for the copied data. The original file remains in Files and its full checksum, byte count, technical metadata, view axes/indices and reviewed plane calibration remain in the authored formula comment.

Admission is at most 1,000,000 finite numeric samples, with an exact 2 MiB UTF-8 bound on the complete authored model source including its existing source and provenance. Small literals can fill that voxel budget; longer literals may reach the byte limit sooner. Unsupported channel layouts or non-rank-3 output, nonfinite selected samples and oversized complete source refuse without sampling, clipping or truncation. The native preparation and existing acknowledged source writer remain the only data and commit paths.

Retained rank-4 TIFF stacks with stored page/channel/row/column order also offer this action. Choose Channel to copy one intensity, RGB, stored alpha or extra-sample channel across every retained page. The rank-3 result removes only the channel axis; native retains the original shape, ordering, channel selection and technical metadata in its provenance and remaps the preview axes. Alpha and intensity samples are copied unchanged. Unsupported planar or transformed TIFF layouts refuse. PNG remains a single image, while NIfTI and DICOM imports are scalar. The original import admission still limits all decoded channels together to 1,000,000 samples, so a four-channel stack can contain at most 250,000 voxels per selected channel.

Choose Use model stack, review the prepared stored axes and spacing, then Save image configuration. Import creates immutable data and no spatial declaration. DICOM frames remain a frame collection in retained provenance. Rank, pages and channels do not establish spatial geometry. For a spatial model volume, separately open Spatial volume declaration, review the depth/row/column meanings and optionally provide positive complete three-axis spacing before whole-volume measurements.

Threshold a reviewed spatial volume

A model image with an explicit reviewed rank-3 spatial declaration offers Volume segmentation → Create volume threshold mask without requiring an existing mask. The native IMAGE_THRESHOLD_3D(source, size0, size1, size2, threshold) labels samples at or above a finite threshold in their original intensity units, preserving the exact stored shape. It checks every voxel and refuses missing/error/string/nonfinite samples rather than silently replacing them or normalizing values. Admission is at most 1,000,000 voxels.

Name the fresh mask in folded options if needed, create it, review the native current-plane preview, then Use volume mask and explicitly save the configuration. The original intensities and masks remain. After connecting the binary mask, label connected volume regions and analyze or retain exact native whole-volume captures. Changed owners, geometry, source or revisions refuse connection; uncertain Code writes retain their draft and are observed without replay.

Scientific JSON downloads, saved measurement records and recovery copies preserve negative zero as a JSON number. Exact save receipts distinguish -0 from 0; CSV numeric fields preserve the same sign. Existing complete-record byte limits apply to the actual serialized bytes.

Volume-box drafts retain their edited footprint while you browse stored planes or inspect pixels. A box stays tied to its source hash, shape, displayed axes and fixed slice/channel indices; browsing its depth axis is allowed. Return to the box plane or explicitly discard it before measuring another plane/channel. Resuming a saved box keeps the original record unchanged while the new draft is reviewed and remeasured.

During a native region or volume measurement, Stop measurement ends the wait and keeps the bounds, outline and label. Measure again when ready; stopping leaves no fresh result available to save. A source or plane reload releases the obsolete request, and a late response cannot replace or unlock a newer measurement.

An uncertain region save can be observed after source or calibration changes within the same saved document and binding owner. A changed source or newer local ROI draft requires explicit Use saved regions and keep draft adoption; observation never clears those edits or repeats the original write. Adoption requires a newer saved revision and keeps retained measurements available for download.

Inspect selected area keeps the chosen pixel selected when its detail window opens, including its exact stored coordinates and freshly returned native value. Source and plane changes still invalidate the target; detail paging clears the selection rather than assigning it to another pixel.

Paged mask and volume summaries include folded Find label controls. Enter an exact native label, including 0 for background, to jump to and highlight its recorded row and missing-intensity details. Lookup does not inspect the current mask; complete summary exports and print retain all labels.