The tutorials use a fair amount of DICOM vocabulary, and explaining each term where it first appears would make them unreadable.
This section answers the question directly instead: what a term means, when it is used, and what to watch out for.
Each entry ends with a link to the page that shows how to work with it in Weasis.
Window and level select which range of stored pixel values is mapped to the shades of gray on screen.
Subsections of Glossary
What Are DICOM RT Objects?
Contours, plan and dose instead of images
Radiotherapy uses a family of DICOM objects that describe a treatment rather than an image.
An RT Structure Set (RTSTRUCT) holds the contours drawn on the planning images: the target volumes and the organs at risk, each as a named set of outlines with a frame of reference tying it to the CT it was drawn on.
An RT Plan describes how the treatment is to be delivered — beams, segments, fractions.
An RT Dose carries the computed dose as a three-dimensional grid, displayed as isodose lines over the anatomy or as a color wash.
The dose-volume histogram (DVH) summarizes that grid against the contours: for each structure, how much of its volume receives at least a given dose.
It is the standard way to check that a plan covers the target while sparing what surrounds it, and it can be stored in the RT Dose object or recomputed from the dose grid and the structures.
Read together, these objects let a viewer show contours, isodoses and the DVH against the planning images — which is a different task from reading a diagnostic study, and why viewers treat them separately.
A DICOM Segmentation, usually written SEG, is a DICOM object that stores which voxels belong to what: one or more labeled segments — an organ, a lesion, a target volume — each referencing the frames of the source series they were derived from.
The segmentation travels as its own file, so the original images are never modified.
Every segment carries coded metadata: an anatomical category and type, and how it was produced — manually, semi-automatically or automatically, usually with the name of the algorithm — which is how a viewer can tell a generated segment from a hand-drawn one.
Three storage types exist.
BINARY marks each voxel as in or out, packed one bit per voxel.
FRACTIONAL stores a value per voxel, expressing either the probability that the voxel belongs to the segment or the fraction of it actually covered — which is what many AI models produce, and what lets a viewer draw a graded overlay rather than a hard edge.
LABELMAP packs many segments into a single image by storing the segment number in the voxel value: compact for an object carrying dozens of structures, but, unlike the other two, with no room for segments that overlap.
Lining a segmentation up with images is not only a matter of the references it declares.
A SEG names the series and the frames it was derived from, but it also shares a frame of reference with them — the same patient coordinate system — so a viewer can place the masks by spatial position when those references are missing or point at another reconstruction of the same acquisition.
Its grid is defined independently of the source resolution, so drawing the overlay may involve resampling.
A segmentation is not an RT Structure Set: RT stores outlines as polygons in patient coordinates, SEG stores voxels, and converting between them loses something either way.
Whichever it is, it is an assertion about the images rather than a correction of them, and it is worth no more than whatever produced it — which is why the algorithm is recorded alongside the segment.
A DICOM Structured Report (SR) stores the content of a report as a tree of coded items instead of a block of prose.
Each item pairs a concept with a value — a measurement and its unit, a finding and its code, a piece of text, a reference to an image — and declares how it relates to the item above it: contains, has properties, inferred from.
The nesting therefore carries meaning of its own, expressing which observation belongs to which finding and what each one was derived from.
What makes the content usable by software is that the concepts and most of the values are coded rather than written out.
They are drawn from controlled vocabularies — DICOM’s own, SNOMED CT for anatomy and findings, UCUM for units — and standard templates, each identified by a TID, fix which items a given kind of report is expected to contain.
A measurement recorded that way can be extracted, compared across examinations and fed into a database without anyone parsing sentences.
Two kinds of item make an SR useful in a viewer.
Image references tie an observation to the exact image it was made on, and spatial coordinates — SCOORD within the plane of an image, SCOORD3D in the frame of reference — tie it to a position, so a viewer can take you to the measurement rather than just telling you it exists.
SR comes in several flavors, from the deliberately limited Basic Text SR to Comprehensive 3D SR, which differ in the value types and references they are allowed to use, alongside specialized ones for CAD findings and radiation dose.
Every SR also records how far it got: a completion flag (partial or complete) and a verification flag, which says whether a responsible observer has signed off on the content — worth a glance before a finding is acted upon.
SR is what carries the measurements produced by CAD and AI tools, ultrasound machines and dose monitoring systems.
What a reader sees is a structured document, not a rendered page: the layout is the viewer’s decision, not the file’s, so two viewers may legitimately present the same SR differently, or surface less of it, depending on the templates they recognize.
A report encapsulated as a PDF is not an SR — it displays, but nothing inside it can be read by software.
DICOMweb is the part of the DICOM standard that defines web services over HTTP, as an alternative to the classic DIMSE network protocol.
Instead of a dedicated port and association negotiation, a client makes ordinary HTTP requests — which is what makes a browser, a portal or a firewalled viewer able to talk to an archive at all.
Three services carry most of the traffic.
QIDO-RS searches for studies, series or instances and answers with metadata, typically JSON.
WADO-RS retrieves — the instances themselves, their metadata, rendered frames or bulk pixel data.
STOW-RS stores, sending instances to the archive.
A fourth, UPS-RS, manages worklist items.
Practically, DICOMweb means an archive is reachable with a URL, standard HTTP status codes and an Authorization header — so a token can travel in a header instead of being pasted into a query string, and access can be brokered by the same infrastructure as the rest of a hospital’s web traffic.
It does not replace DIMSE everywhere: C-FIND, C-GET, C-MOVE and C-STORE remain how most archives and modalities talk to each other inside the network.
Maximum intensity projection (MIP) casts a ray through a slab of the volume for every pixel of the output image and keeps the highest value it meets.
Bright structures therefore survive and everything dimmer disappears.
That makes it the natural way to look at anything bright and thin spread over many slices: contrast-filled vessels, calcifications, a lung nodule that is hard to spot slice by slice.
The trade-off is depth.
Because only the maximum along each ray is kept, a MIP says nothing about which of two bright structures lies in front of the other, and a single very bright voxel — metal, a clip, contrast in an adjacent vessel — can hide what is behind it.
This is why a thin slab is usually more informative than projecting the whole volume: it limits how much anatomy is flattened into one image.
A MIP is a projection, not a rendering: it has no lighting and no surfaces, and it is not a substitute for reading the source slices.
Re-slicing a volume in a plane it was never scanned in
Multiplanar reconstruction (MPR) takes a stack of parallel images — typically axial CT or MR slices — treats it as a volume, and re-slices that volume in another plane.
From one axial acquisition you get coronal and sagittal views, and any oblique plane in between, without scanning the patient again.
It is useful whenever anatomy does not follow the plane of acquisition: a vessel running head-to-foot, a fracture line crossing slices, a structure easier to follow along its own axis than across it.
Two things decide how good the result looks.
Slice spacing — a volume reconstructed from thick, widely spaced slices is blurry or stepped in the other planes, because the data simply is not there.
And geometry — the slices must be parallel, regularly spaced and share a frame of reference; irregular spacing, non-parallel slices or a tilted gantry distort the reconstruction unless the viewer corrects for it.
A reconstructed plane is a computed image, not an acquired one.
Measurements taken on it are only as trustworthy as the geometry underneath.
PET activity normalized by injected dose and body size
The standardized uptake value (SUV) turns the raw activity concentration a PET scanner measures into a number that can be compared between patients and between examinations.
It divides the measured concentration by the injected dose and normalizes for body size — by weight for SUVbw, the most common form, and by lean body mass or body surface area for the variants.
A region of interest usually reports several figures: SUVmax, the single hottest voxel, robust to how the region was drawn but sensitive to noise; SUVmean, the average over the region, steadier but dependent on where the boundary was placed; and SUVpeak as a compromise.
SUV is only as good as the data behind it.
It depends on the injected activity, the injection time and the decay correction being recorded correctly in the DICOM attributes, on the uptake interval between injection and acquisition, and on the reconstruction the scanner applied.
Values from different scanners, protocols or uptake times are not directly comparable, which is why a change in SUV on a follow-up study is read with the acquisition parameters in hand rather than on its own.
A DICOM image stores far more distinct values than a display can show or an eye can separate — CT, for instance, covers thousands of Hounsfield units.
Windowing chooses which part of that range becomes visible: the window width is how wide a band of values is mapped to the available shades of gray, and the window center (the level) is where that band sits.
Narrow the window and small differences become visible while everything outside the band turns pure black or white; widen it and more of the range fits on screen with less contrast between nearby values.
This is why the same CT is read at one setting for lung and another for bone or soft tissue: the pixel data does not change, only the mapping.
The values are not arbitrary.
A study often carries VOI LUT information — either a window width and center chosen at acquisition, or a full lookup table — and viewers add presets per modality so a reader can jump between the settings that matter for a body region.
Because windowing is display only, it never alters the stored pixels; what it does alter is what a screenshot or an exported picture will show.
