Evaluate a DICOM-SEG segmentation in Python¶

A DICOM-SEG (DICOM Segmentation object) is what clinical and research pipelines actually produce. segauge reads it directly, so you can evaluate it as-is without converting to NIfTI first (which can lose or alter the geometry).

pip install "segauge[dicom]"

import segauge as sg

# read a DICOM-SEG to a boolean mask plus its physical voxel spacing
mask, spacing = sg.load_mask("segmentation.dcm")
print(mask.shape, spacing)

# pick a specific segment (DICOM-SEG can hold several)
mask, spacing = sg.load_mask("segmentation.dcm", segment_number=2)

Evaluate prediction vs ground truth, both DICOM-SEG¶

import segauge as sg

result = sg.evaluate([
    sg.Case("patient_001", pred="pred_seg.dcm", gt="gt_seg.dcm"),
])
print(result.summary())        # Dice, HD95, ASSD, NSD, detection F1, all with CIs
result.to_html("report.html")

The voxel spacing is read from the DICOM-SEG itself, so distance metrics (HD95, ASSD, NSD) are computed in real physical units.

RTSTRUCT¶

For RTSTRUCT (contours, common in radiotherapy), point segauge at the structure set and its referenced image series:

mask, spacing = sg.load_rtstruct("ct_series/", "structures.dcm", roi_name="GTV")

Slice spacing is computed from the actual slice positions (ImagePositionPatient), not the SpacingBetweenSlices tag, which is often stale or missing.

See also: Quickstart, the API reference.