One DICOM series in. Read for real by the browser, then preflighted, cleared to leave, reconstructed, or catalogued.

The files never do. Dropped .dcm files are read with FileReader in your own tab, parsed there, and turned into a header dump; the pixel data is never even decoded, because the reader stops at (7FE0,0010). Parsing, the series grouping, the geometry, the identifier audit and all four exports are free, offline and need no account. When you run an AI lane, what is sent is the header text in the box plus the facts the browser measured — and the values of identifying tags are deliberately stripped out of those facts. The model is told that PatientName is present, what its VR is and how many characters it holds, never what it says.

It parses it. A dropped file goes through a DICOM Part 10 reader written for this app: 128-byte preamble, DICM magic, the file meta group read as Explicit VR Little Endian no matter what (0002,0010) then declares the dataset to be, then the dataset in implicit VR, explicit VR little-endian or explicit VR big-endian. The 32-bit-length VRs get their two reserved bytes. Undefined lengths run to their delimiter. Sequences are entered to a bounded depth, so a PersonName nested inside a content sequence is still found. A file with no preamble is recognised as a raw dataset rather than rejected, and a deflated dataset is reported as unreadable rather than silently returned empty.

Because SliceThickness and SpacingBetweenSlices are claims about the acquisition, and the positions in (0020,0032) are the measurement. The app takes the two direction cosines in (0020,0037), crosses them to get the slice normal, and projects every slice's position onto it. The differences between consecutive projections are the spacing — and where they read 2.5, 2.5, 5.0, 2.5 there is a slice missing, whatever the header field says. That is why a gap is reported as a gap rather than as a header value you could have read yourself.

Preflight asks whether the series can be used as it stands, across eight areas: object type, transfer syntax, series separation, geometry, instance completeness, intensity calibration, acquisition consistency and identifiers. De-identify produces a tag-by-tag plan against the PS3.15 profile for the intent you picked, with the residual risks and a pydicom script that writes to a separate directory and maps UIDs consistently. Volume produces the reconstruction plan: what to sort on and why not InstanceNumber, where every geometry number comes from, how the affine is built, and what breaks if it is done naively. Manifest designs the research dataset's columns, their units, the derived fields, the quality columns and the CSV header. All four run over the same header, and each result offers the next lane as a button.

The browser measures first and sends its facts into the run: the series grouping, every geometry number, the identifier inventory and every flag it raised. The model must return one reconciliation entry per critical and high flag, so a defect it ignored is displayed as unaccounted for rather than quietly dropped. Every tag, series and instance name it cites is checked against the header you pasted, and one that does not exist is shown, marked, rather than trusted. Every DICOM tag in the script it returns is checked the same way. On the de-identification lane, any identifying tag the browser found that the plan does not mention is listed as a hole in the plan.

No, and deliberately not. DICOM Desk never rewrites a file. It reads headers, tells you what is in them, and produces a plan and a script for you to run and review yourself. The PS3.15 Annex E action codes it uses are a transcription of one published baseline profile — not legal advice, not a compliance certificate, and not a substitute for your study's own data-use agreement. Nothing here can find an identifier burned into the pixels, either; the app tells you when that risk is unresolved rather than pretending to have checked.

A worst-case amount is reserved before you run and only what the run actually uses is charged. The estimate is checked against your balance before the button enables: signed out, it says signing in comes first; under the model's minimum, it names the exact shortfall; and a reply cut short by a low balance says so rather than presenting itself as complete. The whole browser side is free and needs no account at all.