dicom-series-to-volume

dicom-series-to-volume is an NVIDIA MedTech research-and-development skill that converts a single CT DICOM series folder into a Hounsfield-Unit-scaled NIfTI (.nii.gz) volume with accompanying geometry evidence. It solves a common medical-imaging preprocessing problem: turning a directory of DICOM slices into a correctly scaled, spatially coherent 3D volume that downstream pipelines (for example segmentation models) can consume. It reads one series, sorts slices by ImagePositionPatient, applies RescaleSlope and RescaleIntercept, builds an affine from orientation and spacing tags, and writes the volume plus a JSON summary.

The skill wraps a documented entrypoint (`scripts/series_to_volume.py`) declared in `skill_manifest.yaml`, and instructs agents to use the wrapper exactly as documented rather than reimplementing it. It emits key fields such as n_slices, series_instance_uid, output shape/spacing/axcodes/affine, HU range, and conversion time, and it is paired with a `dicom_volume_quality_v1` verifier and an output schema for validation. A trusted-run path via the eval engine is also provided. It ships with fixtures, benchmarks (NVSkills-Eval scored 100% on the Security dimension), a skill card, and an output validator.

It is intended for developers and engineers building medical-imaging pipelines who need HU-scaled NIfTI volumes with orientation evidence. The scope is deliberately narrow: single-series CT only. It rejects multi-series input at preflight and does not support multi-frame DICOM, compressed transfer syntaxes (JPEG/JPEG2000/RLE), RT structure sets, or auto-reorientation. It is explicitly not for clinical deployment, autonomous diagnosis, regulatory submission, or production inference, for which a vetted converter such as dcm2niix is recommended.