Locating fractional DICOM Segmentations in IDC

Support
1

This post is in response to this question that I received from an IDC user recently:

“where I can find an example of FRACTIONAL Segmentation in DICOM-SEG?”

TL;DR: IDC contains 17,238 fractional segmentation series across multiple radiology and slide microscopy collections.

What are Fractional Segmentations?

DICOM Segmentation objects support three encoding types, defined by the Segmentation Type attribute (0062,0001):

Type Values Overlap Allowed Description
BINARY 0 or 1 Yes Each voxel is fully inside (1) or outside (0) the segment
FRACTIONAL 0 to MaxFractionalValue Yes Continuous values representing probability or occupancy
LABELMAP Integer segment labels No Each voxel belongs to exactly one segment (mutually exclusive)

Fractional Sub-types

FRACTIONAL segmentations are further classified by Segmentation Fractional Type (0062,0010):

  • PROBABILITY: Value represents the probability that the segment occupies the voxel
  • OCCUPANCY: Value represents the proportion of voxel volume occupied by the segment

Values are stored as unsigned 8-bit integers with the maximum value defined by Maximum Fractional Value attribute (0062,000E) (up to 255).

Why Use Fractional Segmentations?

Fractional segmentations are useful for representing:

  • Probability maps from AI/ML model predictions
  • Partial volume effects where a voxel contains multiple tissue types
  • Uncertainty quantification in automated segmentations
  • Continuous measurements like enhancement ratios in functional imaging

Note: as of data release v23, IDC contains only BINARY and FRACTIONAL segmentations; no LABELMAP examples are available.

See DICOM PS3.3 Section C.8.20.2.3 for the normative definition.

Major Collections with Fractional Segmentations

collection_id Source Modality Algorithm Type Count
ispy2 MR Background Threshold, PE threshold SEMIAUTOMATIC 2,688
ispy1 MR Background Threshold variants SEMIAUTOMATIC 2,568
acrin_6698 MR Background Threshold / Manual SEMI/MANUAL 2,213
tcga_brca SM Stony Brook TIL Segmentation AUTOMATIC 1,061
breast_mri_nact_pilot MR Background Threshold, VOI limits SEMIAUTOMATIC 756
tcga_kirc SM Stony Brook TIL Segmentation AUTOMATIC 514
tcga_ucec SM Stony Brook TIL Segmentation AUTOMATIC 504
tcga_luad SM Stony Brook TIL Segmentation AUTOMATIC 479

Two Main Types of Fractional SEGs

  1. Breast MRI (MR) functional tumor mapping (ISPY1, ISPY2, ACRIN-6698) - Probability maps from DCE-MRI enhancement thresholds.
  2. Slide Microscopy (SM) TIL segmentations (TCGA collections) - Deep learning-based predictions on pathology slides

Ready-to-View Examples

Example 1: Breast MRI (Source Modality: MR)

  • collection_id: breast_mri_nact_pilot (part of the I-SPY 1 trial imaging data)
  • Patient: UCSF-BR-26
  • Algorithm: Background Threshold, PEthresh and MinConn filter, VOI limits
  • Source Modality: MR (Breast)
  • License: CC BY 3.0
  • View DICOM study

Note: OHIF v3 currently cannot display fractional DICOM SEG, see handle fractional SEG objects · Issue #1346 · OHIF/Viewers · GitHub.

Example 2: Pathology Slide (Source Modality: SM)

  • collection_id: tcga_kirc
  • Patient: TCGA-CJ-4875
  • Algorithm: Stony Brook TIL Segmentation Inception-V4 2022
  • Source Modality: SM (Slide Microscopy)
  • License: CC BY 4.0
  • View DICOM study

Tip: Generate viewer URLs programmatically with client.get_viewer_URL(seriesInstanceUID="...") which auto-selects OHIF for radiology or SLIM for pathology.

Download an Example

from idc_index import IDCClient
client = IDCClient()

# Download a fractional SEG from breast MRI (Source Modality: MR)
# Collection: breast_mri_nact_pilot, Patient: UCSF-BR-26
client.download_from_selection(
    seriesInstanceUID="1.3.6.1.4.1.14519.5.2.1.7695.2311.249901144986078351313294733363",
    downloadDir="./fractional_seg_example_MR"
)

# Download a fractional SEG from pathology slide (Source Modality: SM)
# Collection: tcga_kirc, Patient: TCGA-CJ-4875
client.download_from_selection(
    seriesInstanceUID="1.2.826.0.1.3680043.10.511.3.43729124912062621812911227783921569",
    downloadDir="./fractional_seg_example_SM"
)

Query for All Fractional Segmentations

from idc_index import IDCClient
client = IDCClient()
client.fetch_index("seg_index")

# Find all fractional segmentations with collection info and source modality
results = client.sql_query("""
    SELECT
        i.collection_id,
        src.Modality as source_modality,
        s.SeriesInstanceUID,
        s.AlgorithmName,
        s.AlgorithmType,
        s.total_segments,
        s.segmented_SeriesInstanceUID as source_series
    FROM seg_index s
    JOIN index i ON s.SeriesInstanceUID = i.SeriesInstanceUID
    JOIN index src ON s.segmented_SeriesInstanceUID = src.SeriesInstanceUID
    WHERE s.SegmentationType = 'FRACTIONAL'
""")
print(f"Found {len(results)} fractional segmentation series")
print(results[['collection_id', 'source_modality', 'AlgorithmName']].head(10))

Verify the Data

Run this to confirm the statistics are current:

from idc_index import IDCClient
client = IDCClient()
client.fetch_index("seg_index")

# Total fractional segmentations
total = client.sql_query("SELECT COUNT(*) FROM seg_index WHERE SegmentationType = 'FRACTIONAL'")
print(f"Total fractional SEGs: {total.iloc[0,0]}")

# IDC version
print(f"IDC version: {client.get_idc_version()}")

If you made it all the way to this point - you get a bonus: this blog post was generated using the imaging-data-commons Claude skill that we recently added here: GitHub - K-Dense-AI/claude-scientific-skills: A set of ready to use scientific skills for Claude (with several follow up prompts, not from the first try yet), followed by some relatively minor manual edits. I will follow up with a separate post on that skill, but you are welcome to give it a try right away and share your feedback with me!