DICOM Segmentation Object

The Segmentation Object supports two distinct encoding types defined by the Segmentation Type attribute (0062,0001):

• BINARY: Each segment is a separate bit-plane in the pixel data, where a value of 1 indicates pixel membership in the region of interest. Multiple segments can be encoded in a single frame using bit-packed storage.
• FRACTIONAL: Each segment occupies an entire frame with grayscale values representing the probability or partial volume of tissue membership, typically ranging from 0 to 255 for 8-bit data.

The choice between binary and fractional encoding directly impacts storage efficiency and the clinical precision of boundary representation.

Unlike single-frame DICOM images, the Segmentation Object is inherently a multi-frame image where each frame represents a distinct segment or slice position:

• Frames are organized using the Dimension Organization module, which defines index sequences for spatial position, segment number, and temporal points.
• The Dimension Index Sequence (0020,9222) explicitly maps each frame to its corresponding segment number and spatial coordinates.
• This structure allows a single SOP Instance to contain hundreds of segments across a full volumetric dataset, dramatically reducing the number of objects that must be managed compared to storing each segmentation as a separate image.

Each segmented region is described by a rich set of metadata using the Segment Sequence (0062,0002):

• Segment Number (0062,0004): Uniquely identifies the segment within the SOP Instance.
• Segment Label (0062,0005): A human-readable name such as "Liver" or "Tumor Core."
• Segmented Property Category Code Sequence (0062,0003): Maps the segment to a coded concept from controlled terminologies like SNOMED CT or FMA, ensuring semantic interoperability.
• Anatomic Region Sequence (0008,2218): Specifies the body part examined, enabling automated routing and hanging protocol selection.
• Tracking ID (0062,0020) and Tracking UID (0062,0021): Support longitudinal tracking of the same anatomical feature across multiple studies.

The Segmentation Object maintains explicit spatial relationships to its source images through two critical mechanisms:

• Derivation Image Sequence (0008,9124): References all source SOP Instances from which the segmentation was derived, establishing provenance and enabling downstream audit trails.
• Spatial Registration via the Registration Sequence (0070,0308) within the Common Instance Reference Module: Defines the rigid or deformable transformation matrix that maps the segmentation's frame of reference to the referenced image's coordinate system.
• The Frame of Reference UID (0020,0052) ensures that the segmentation shares the same spatial coordinate space as the source images, critical for multi-modality fusion.

Beyond pixel-based encoding, the Segmentation Object can store surface representations using the Surface Segmentation Module:

• Surface Sequence (0066,0002): Contains one or more surface meshes, each defined by vertex coordinates and triangle connectivity.
• Surface Points Sequence (0066,0011) and Surface Faces Sequence (0066,0027): Store the raw geometry data.
• Recommended Display Grayscale Value (0062,000C): Specifies the preferred display intensity for rendering the surface.
• Surface meshes are particularly valuable for surgical planning and 3D printing workflows, where explicit boundary geometry is required rather than voxel masks.

The DICOM standard defines a specific SOP Class UID for the Segmentation Object:

• SOP Class UID: 1.2.840.10008.5.1.4.1.1.66.4
• This UID is negotiated during Association Negotiation between an SCU and SCP to confirm mutual support for segmentation storage.
• The corresponding Storage SOP Class is formally named "Segmentation Storage."
• When stored as a DICOM Part 10 file, the Media Storage SOP Class UID in the File Meta Information header must match this identifier to ensure correct parsing by any standards-compliant PACS or VNA.

The DICOM Segmentation Object supports binary (0 or 1) and fractional (probability maps) segmentation types. Binary segmentation uses a single bit per pixel to define a hard mask, while fractional encoding stores a continuous value between 0 and 1 per pixel, representing the probability of tissue membership. This is critical for storing the raw output of deep learning models before thresholding, preserving uncertainty information for downstream quantitative analysis.

A Segmentation Object must explicitly reference the source DICOM images from which it was derived via the Referenced Series Sequence and Referenced SOP Instance UID. The spatial registration between the segmentation frames and the source images is defined using a rigid transformation matrix stored in the Registration Sequence or via identical Frame of Reference UID and Image Position/Orientation attributes, ensuring pixel-perfect overlay in a PACS viewer.

Each distinct region within a Segmentation Object is assigned a unique Segment Number (integer, starting at 1) and can be labeled with a coded anatomic concept from a controlled terminology such as SNOMED CT or RadLex. The Segmented Property Category Code Sequence and Segmented Property Type Code Sequence allow the object to self-describe that Segment 1 represents 'Lung' and Segment 2 represents 'Tumor', enabling automated ingestion by radiomics platforms.

While both encode regions of interest, the DICOM Segmentation Object stores results as a pixel-aligned raster mask, whereas the RT Structure Set stores contours as geometric planar coordinates. Key distinctions:

• Segmentation Object: Ideal for AI-generated voxel-wise output; supports fractional encoding; natively aligned to image grid.
• RT Structure Set: Ideal for human-drawn contours; supports non-axial planes; required for radiotherapy treatment planning systems. Conversion between the two formats is a common interoperability challenge.

A single DICOM Segmentation Object is a multi-frame SOP Instance where each frame corresponds to a single slice of a 3D volume. The Per-Frame Functional Groups Sequence allows each frame to carry its own spatial metadata, including Plane Position and Plane Orientation, enabling the segmentation to be correctly mapped to a non-uniform or tilted source acquisition. This structure efficiently packages an entire 3D organ mask into one manageable DICOM object.

Beyond raster masks, the DICOM standard also defines a Surface Segmentation SOP Class for storing 3D surface meshes as triangulated point clouds. This is particularly useful for surgical planning and 3D printing applications where a lightweight geometric boundary representation is preferred over a dense volumetric mask. The surface mesh is stored using the Surface Mesh Module, which encodes vertex coordinates and triangle connectivity.