wenmin-wu/cv-dicom-ipp-z-slice-sort

Overview

DICOM filenames in a CT series are not guaranteed to be in anatomical order. Different scanners, anonymizers, and export tools produce wildly inconsistent name schemes — 1.dcm may be the topmost slice on one study and the bottommost on the next. InstanceNumber is also unreliable: it can be reset, missing, or assigned in scan order rather than spatial order. The single trustworthy ordering is ImagePositionPatient[2], the Z coordinate of the slice in patient space — sort by it ascending and you get a consistent head→feet (or feet→head, fix later) volume regardless of vendor. Keep an integer-filename fallback for the few exports that strip the tag entirely.

Quick Start

import pydicom
from os import listdir

def load_scans(dcm_dir):
    files = listdir(dcm_dir)
    try:
        slices = [pydicom.dcmread(f'{dcm_dir}/{f}') for f in files]
        slices.sort(key=lambda s: float(s.ImagePositionPatient[2]))
    except (AttributeError, KeyError):
        # fallback: integer-filename order when IPP tag is missing
        nums = sorted(int(f.split('.')[0]) for f in files)
        slices = [pydicom.dcmread(f'{dcm_dir}/{n}.dcm') for n in nums]
    return slices

Workflow

1. List every .dcm file in the series directory
2. Read each with pydicom.dcmread and key the sort on float(ds.ImagePositionPatient[2])
3. Wrap in try/except (AttributeError, KeyError) to catch the rare missing-tag series
4. In the fallback branch, sort by the integer extracted from the filename stem
5. Stack pixel_arrays in the resulting order and apply RescaleSlope/Intercept to get HU

Key Decisions

• ImagePositionPatient[2], not [0] or [1]: index 2 is the through-plane axis (Z) for axial CT; the other indices are in-plane translation and are constant across slices.
• Sort ascending, flip later: produces deterministic ordering; canonical head→feet flip can be done as a separate step (see cv-ct-z-stack-orientation-flip).
• Integer-filename fallback, not lexicographic: '10.dcm' sorts before '2.dcm' lexicographically; integer parsing fixes it.
• float() cast required: pydicom returns the position as DSfloat which is a string-ish type and won't compare correctly without the cast.
• Keep the slice objects together with the position: don't reorder a separate position array independently of the pixel arrays — index drift bugs are silent and brutal.

References

• Pulmonary Dicom Preprocessing