wenmin-wu/cv-ct-z-stack-orientation-flip

Overview

DICOM file names are not a reliable proxy for slice order. The same scanner can write 1.dcm as the topmost slice on one acquisition and the bottommost on the next, depending on protocol and reconstruction direction. If you train on a mix and don't normalize, augmentations like flip(axis=0) become silently inconsistent and the model learns a much fuzzier through-plane signal than it could. The single-line fix: read ImagePositionPatient[2] from the first and last DICOM in the series; if the last has a larger Z value than the first, the stack is inverted relative to the canonical patient frame and you flip it.

Quick Start

import dicomsdl
import numpy as np

def canonicalize_z_order(image, dcm_dir, slice_min, slice_max):
    dcm0 = dicomsdl.open(f'{dcm_dir}/{slice_min}.dcm')
    dcmN = dicomsdl.open(f'{dcm_dir}/{slice_max - 1}.dcm')
    z0 = dcm0.ImagePositionPatient[2]
    zN = dcmN.ImagePositionPatient[2]

    if zN > z0:                       # inverted: flip into head→feet order
        image = image[::-1]

    dz = abs((zN - z0) / max(slice_max - slice_min - 1, 1))
    return np.ascontiguousarray(image), dz

Workflow

1. After loading every slice into a (D, H, W) array (in filename order), open just the first and last DICOM headers
2. Compare ImagePositionPatient[2] — the third element is the Z coordinate in patient space
3. If zN > z0, the stack is in feet→head order; reverse it with image[::-1]
4. Compute dz as the absolute difference divided by (num_slices - 1) and store it for the resampling step
5. np.ascontiguousarray after the slice-reverse to avoid downstream RuntimeError: non-contiguous from torch

Key Decisions

• Compare Z, not InstanceNumber: InstanceNumber can be reset, missing, or inconsistent across vendors; ImagePositionPatient is the geometric ground truth.
• Absolute dz: after flipping, the sign is meaningless — what the resampler needs is the magnitude.
• Only open two DICOMs, not all of them: the orientation check costs ~1ms and avoids re-parsing every header.
• np.ascontiguousarray after [::-1]: numpy reverse-views are non-contiguous and break torch tensor zero-copy paths.
• Don't sort by Z to "fix" ordering: just flipping is faster and equivalent for evenly-spaced acquisitions, which is what 99% of CT series are.

References

• LB 0.55 — 2.5D + 3D sample model