segauge

DICOM-native evaluation for medical image segmentation: surface-mesh distance metrics, per-lesion detection, fairness slices, and a confidence interval on every number.

segauge computes Dice, IoU, Hausdorff distance (HD), HD95, average symmetric surface distance (ASSD), Normalized Surface Dice (NSD), and per-lesion detection F1 for 3D medical image segmentation, and puts a bootstrap confidence interval on every number. It reads NIfTI, DICOM-SEG, RTSTRUCT, and NumPy directly, so you can evaluate the output of nnU-Net, MONAI, or TotalSegmentator without a lossy conversion.

It exists because the metrics you report should be the metrics that are true, and today they often aren't.

pip install segauge

import segauge as sg

result = sg.evaluate([
    sg.Case("patient_001", pred="pred.nii.gz", gt="gt.nii.gz", metadata={"scanner": "siemens"}),
    sg.Case("patient_002", pred="pred2.dcm",   gt="gt2.dcm",   metadata={"scanner": "ge"}),
])

print(result.summary())          # every metric with a 95% CI
print(result.by_subgroup("scanner"))   # where does the model quietly fail?
result.to_html("report.html")    # one self-contained report

Or from the command line:

segauge eval --pred preds/ --gt labels/ --metadata cases.csv --report report.html

Why not just use MONAI or Metrics Reloaded?

segauge bundles four things no incumbent provides together: a confidence interval on every metric, per-lesion detection, subgroup/fairness slicing, and native DICOM-SEG/RTSTRUCT input.

It also computes distance metrics on a surface mesh rather than the voxel grid: it extracts the object surface with marching cubes at true voxel spacing and integrates over the surface, following the MeshMetrics method (Podobnik & Vrtovec, 2025). In our own benchmark (benchmarks/mesh_vs_grid.py) this measurably reduces HD95 error under anisotropic, thick-slice spacing, where grid rasterization hurts most; for mean distance (ASSD) on smooth shapes the two methods are comparable. A formal validation suite against the MeshMetrics reference is in progress; we report what the benchmark shows, not more.

Metrics

• Overlap: Dice (DSC), IoU (Jaccard) — exact, integer-counted.
• Surface distance: Hausdorff (HD), HD95, ASSD, MASD, Normalized Surface Dice (NSD) — mesh-based, spacing-aware, area-weighted.
• Per-lesion detection: precision, recall, F1 via connected-component matching, so you can answer "did it find the tumor?" not just "how much voxel overlap?"
• Every aggregate carries a deterministic bootstrap confidence interval.

Inputs

NIfTI (.nii, .nii.gz), DICOM-SEG, RTSTRUCT, NumPy arrays, and .npy. Voxel spacing is read from the file header and used for the distance metrics, so a segmentation from a clinical pipeline is evaluated as-is.

FAQ

How do I compute HD95 correctly in Python? segauge.surface_metrics(pred, gt, spacing) returns HD, HD95, ASSD, MASD, and NSD computed on the surface mesh, not the voxel grid.

Can it evaluate DICOM-SEG / RTSTRUCT directly? Yes. pip install segauge[dicom] and pass a .dcm SEG, or use segauge.load_rtstruct(series_dir, rtstruct, roi_name).

Does it work with nnU-Net / TotalSegmentator / MONAI outputs? Yes. Point it at the NIfTI (or DICOM) they produce.

Why a confidence interval? A Dice of 0.85 on 12 cases is not the same claim as 0.85 on 1200. segauge makes the difference visible.

2D images? Overlap and detection metrics work in any dimension; surface-distance metrics are 3D in v0.1 (2D is planned for v0.2).

Status

Pre-release (0.1.0.dev), built in the open. segauge is an evaluation tool for developers and researchers. It is not a medical device and produces no diagnosis.

References

• Maier-Hein, Reinke et al. Metrics Reloaded. Nature Methods (2024). https://www.nature.com/articles/s41592-023-02151-z
• Podobnik & Vrtovec. MeshMetrics. arXiv:2509.05670 (2025). https://arxiv.org/abs/2509.05670

License

Apache-2.0.