Segmentation Types¶
Multi-class Segmentations¶
In the case of multi-class segmentations, a pixel or voxel can only belong to a single class. Thus, all classes depend on each other. This kind of segmentation is also sometimes called label map. One possible application for multi-class segmentation would be abdominal organ segmentation in CT or MR images. Each pixel/voxel can only be assigned to a single organ, but with multiple organs present in the image.
Example¶
DICOM Data Elements¶
Tag |
Tag Name |
Value |
|---|---|---|
(0x0062,0x0001) |
SegmentationType |
BINARY |
(0x0062,0x0013) |
SegmentsOverlap |
NO |
(0x0028,0x0100) |
BitsAllocated |
1 |
(0x0028,0x0101) |
BitsStored |
1 |
(0x0028,0x0102) |
HighBit |
0 |
Multi-label Segmentations¶
The main difference between multi-class and multi-label segmentation is the independence of classes. A pixel/voxel can belong to more than a single class. For instance, one encoded segment in the DICOM-SEG could represent a liver, while the next segment would represent tumors within the liver. In this case, both segments could even build a semantic relationship. Maybe an annotator segments tumors in the abdomen, but does not differentiate between different organs. Additionally, multiple organs are fully annotated and thus the tumor location can be derived from the combination of classes.
Example¶
DICOM Data Elements¶
Tag |
Tag Name |
Value |
|---|---|---|
(0x0062,0x0001) |
SegmentationType |
BINARY |
(0x0062,0x0013) |
SegmentsOverlap |
YES or UNDEFINED |
(0x0028,0x0100) |
BitsAllocated |
1 |
(0x0028,0x0101) |
BitsStored |
1 |
(0x0028,0x0102) |
HighBit |
0 |
Fractional Segmentations¶
The DICOM-SEG standard allows for fractional segmentations, e.g. for representing probabilities or the ratio of tissue composition. A floating point value between 0.0 and 1.0 will be quantized into an unsigned integer with 8-bits, which results in a minimum step size of ~0.39% between two integer values. Possible use cases are:
Class probabilities of semantic segmentation algorithms
Uncertainty estimations of semantic segmentation algorithms
Material specific decomposition in Dual-Energy Computed Tomography (DECT)
DICOM Data Elements¶
Tag |
Tag Name |
Value |
|---|---|---|
(0x0062,0x0001) |
SegmentationType |
FRACTIONAL |
(0x0062,0x0010) |
SegmentationFractionalType |
PROBABILITY or OCCUPANCY |
(0x0062,0x000e) |
MaximumFractionalValue |
255 |
(0x0028,0x0100) |
BitsAllocated |
8 |
(0x0028,0x0101) |
BitsStored |
1 |
(0x0028,0x0102) |
HighBit |
0 |