Enhanced Multi-Frame DICOM: Functional Groups Explained

  • Enhanced multi-frame DICOM stores an entire scan as one instance instead of one file per slice.
  • Functional groups live in the Shared Functional Groups Sequence or the Per-Frame Functional Groups Sequence.
  • A functional group appears in only one of those two sequences, never both, for a given attribute.
  • Image Position and Pixel Spacing may live per frame rather than at the top level, unlike classic single-slice files.
  • Large frame sets can split across multiple instances sharing a Concatenation UID and an In-concatenation Number.

An enhanced multi-frame DICOM object is one file that holds many image frames, plus metadata that describes each frame. Modern CT, MR, and PET scanners use it so a scan can be one instance instead of a thousand. If your product reads imaging data, you will meet it sooner or later. The good news: it is fully defined by the public standard, and the rules are regular once you see the pattern.

What is a multi-frame DICOM image?

A multi-frame image is a DICOM image whose pixel data holds a sequence of frames. The standard says this in PS3.3 C.7.6.6. Frames travel as one continuous stream of pixels, and no frame headers exist inside it.

Number of Frames (0028,0008) tells you how many frames there are. It must be greater than zero. A single-frame image is just the simple case.

Classic CT and MR work differently. Each slice is its own instance with its own SOP Instance UID. If you have not read how that works, start with instances, series, and studies. Multi-frame changes the unit of storage, not the idea of a study.

What makes an “enhanced” multi-frame object different?

Enhanced IODs add Functional Groups to multi-frame images. A Functional Group is a set of attributes that belong together and may change from frame to frame. Enhanced CT, Enhanced MR, and Enhanced PET all use them.

Multi-frame is the pixel layout. Enhanced is the metadata layout on top of it. Frame Increment Pointer (0028,0009) only names the attribute that orders the frames, for example Frame Time on a cine loop. Functional groups add position, spacing, and value transforms for each frame.

Enhanced objects carry the Multi-frame Functional Groups Module (PS3.3 C.7.6.16). The standard says this module is present even if the instance holds only one frame.

Enhanced SOP Classes have their own UIDs. Two examples from PS3.4 Annex B:

  • Enhanced CT Image Storage: 1.2.840.10008.5.1.4.1.1.2.1
  • Enhanced MR Image Storage: 1.2.840.10008.5.1.4.1.1.4.1

The same table lists Legacy Converted Enhanced CT Image Storage, 1.2.840.10008.5.1.4.1.1.2.2, as a separate SOP Class with its own UID. Classic CT Image Storage (1.2.840.10008.5.1.4.1.1.2) still exists too. “Enhanced” describes the metadata layout, not better image quality. A receiver that supports only the classic SOP Classes will not accept the enhanced ones, so read the sender’s conformance statement first. It must list the functional groups the sender can include.

How do functional groups describe each frame?

Functional groups live in two sequences. The Shared Functional Groups Sequence (5200,9229) holds values that are the same for every frame. The Per-Frame Functional Groups Sequence (5200,9230) holds one item per frame, in frame order. The first item belongs to the first frame, and so on.

A group appears in one place only. The standard does not allow the same functional group in both sequences.

This design saves space. Pixel spacing is usually the same for a whole volume, so it sits once in the shared sequence. Slice position changes with every frame, so it sits in each per-frame item.

These are the groups you will use most:

What you need Sequence Usually
Pixel Spacing, Slice Thickness Pixel Measures Sequence (0028,9110) Shared
Image Position (Patient) Plane Position Sequence (0020,9113) Per frame
Image Orientation (Patient) Plane Orientation Sequence (0020,9116) Shared
Rescale Slope and Intercept Pixel Value Transformation Sequence (0028,9145) Shared or per frame
Dimension indexes Frame Content Sequence (0020,9111) Per frame

“Usually” is the key word. An IOD allows both places for some groups, and the scanner chooses. Your code must look in the per-frame item first, then fall back to the shared item. Reading only one of them is the most common bug with enhanced files.

Another detail: when per-frame data exists, the standard requires one item for every frame, even if that item is empty. Check that the item count equals Number of Frames before you trust the data.

How do you read an enhanced multi-frame file?

Treat the frame index as the key and rebuild each frame’s metadata from the two sequences. The steps are short:

  1. Read Number of Frames and the Rows, Columns, and Bits Allocated values.
  2. Split the pixel data into frames. For native pixel data, each frame is Rows × Columns × samples × bytes per sample.
  3. For frame i, read per-frame item i, then merge in the shared item.
  4. Sort frames by the position or by the dimension index values you need.

If you worked through sorting slices by Image Position, step 4 will look familiar. The difference is where the numbers come from. Single-slice files hold position at the top level. Enhanced files hold it inside the Plane Position Sequence of each frame.

Be careful with attributes at the top level. The standard says all Image IE attributes relate to the first frame. A top-level value is not a safe description of frame 50.

What changes for storage, viewers, and ML pipelines?

One file with many frames is a different size problem from many small files. You get one SOP Instance UID, one object, and all frames move together.

Storage. Your instance count drops, but each object is large. One file is one row in your data model, so add a frame-count column and keep the per-frame metadata inside the original file. Split it into rows only if a feature needs it. Any layer that sets a request-body limit must allow your largest single instance, as in large DICOM studies.

Viewers. A viewer needs frame-level access. It should not decode 1,000 frames to show frame 500. With compressed data, DICOM stores each frame in encapsulated fragments (PS3.5 A.4). A Basic Offset Table at the start can list the byte offset of each frame’s first fragment. It is not required to have a value, so decoders must handle both cases. DICOMweb also defines a Frame Pixel Data resource, /studies/{study}/series/{series}/instances/{instance}/frames/{frames}, so a client can ask a server for chosen frames rather than the whole instance (PS3.18 10.4).

ML data loaders. Loading every frame into memory at once costs a lot of memory, so read by frame when you can. If you train on volumes, one enhanced file may already hold the whole volume, which is convenient. Old code that reads Image Position (Patient) from the top level will find nothing, because it now sits inside the Plane Position Sequence. Read Rescale Slope and Intercept from the Pixel Value Transformation group and spacing from Pixel Measures, as in DICOM pixel data for AI and ML engineers. Compression choices change decode cost, which DICOM compression transfer syntaxes compared covers.

What limits and pitfalls should you plan for?

One limit is real: concatenation. A very large frame set may be split across several instances that share a Concatenation UID, with In-concatenation Number giving the order. The standard describes this in the Multi-frame Functional Groups Module. If your product assumes “one instance equals one acquisition”, it will miss data when it meets a concatenation.

Plan for these other cases too:

  • Old and new together. Your code may see classic slice files and enhanced files in the same study list. Support both, and test with files from more than one vendor.
  • Frame numbers. DICOM numbers frames from 1. Many libraries use zero-based indexes, so convert in one place.
  • Private data. Vendors still add private attributes inside functional groups. Read how to handle private tags before you decide what to keep.

None of these is hard once you know it exists. DICOM is a stable, free, public standard, and enhanced objects are well documented. A product that reads them correctly, whether a viewer, an AI pipeline, or a research platform, handles something many tools still get wrong.

What should you do next?

Open one enhanced CT or MR file from a public sample set and print its Shared and Per-Frame Functional Groups Sequences. Check that the per-frame item count equals Number of Frames, then find where Image Position (Patient) and Pixel Spacing live. That one exercise teaches more than any summary.

Use synthetic test data, never real patient files. Then read PS3.3 C.7.6.16 once, top to bottom. It is short, and it is the exact contract your parser has to follow.

Dicomly builds infrastructure for healthcare software teams that receive DICOM studies from hospitals. Why we built Dicomly →

Sources

  1. DICOM PS3.3 C.7.6.6 — Multi-frame Module (dicom.nema.org)
  2. DICOM PS3.3 C.7.6.16 — Multi-frame Functional Groups Module (dicom.nema.org)
  3. DICOM PS3.5 A.4 — Transfer Syntaxes for Encapsulation of Encoded Pixel Data (dicom.nema.org)
  4. DICOM PS3.4 B.5 — Standard SOP Classes (dicom.nema.org)

Start receiving DICOM today.

Early access. No credit card. First endpoint free.