ct_mri_dicom_nii_reader

A small, dependency-light Python library for reading medical image volumes (CT / MRI / masks) from DICOM series and NIfTI files into isotropic, LPS-ordered NumPy arrays, plus a compact container class with slice access, processing helpers and optional Tkinter/Pillow preview GUIs.

The classes are extracted verbatim from the regknee knee-registration project so they can be reused by future projects as a self-contained data layer. No command-line interface is provided; this is a pure Python programming interface.

Features

• DICOM series -> isotropic HU volume: load_dicom_hu_lps reads a DICOM directory (auto-selecting the largest multi-slice series), applies Rescale Slope / Intercept, and resamples onto an isotropic LPS grid (float32 array with shape (n_L, n_P, n_S)).
• NIfTI -> isotropic LPS volume: load_nifti_lps reads .nii / .nii.gz files (including non-ASCII paths and mismatched compression on Windows) with conservative CT/MR modality detection.
• BodyData container: keeps a volume together with its voxel size (mmpd) and image type (ct / mri / mask), provides slice access, cloning, percentile queries, ROI extraction, MIND descriptor computation and GUI preview.
• Loader abstraction: BodyDataLoaderManager dispatches by file extension so one call loads DICOM, NIfTI or the compact unified .ubd.npz format.
• Compact .ubd.npz storage: BodyData.save / UnifiedBodyDataLoader round-trip a volume, its type and voxel size through a single compressed NumPy archive.
• Optional GUIs: 2-D and 3-D grayscale preview windows built on tkinter and Pillow (kept for interactive exploration; headless workflows never need them). Windows open at a default upscaled size (smallest image side 512 px); resizing the window rescales the image with it while preserving the original aspect ratio.
• Visualization classes: BodyDataCommonGrid (origin alignment + unified sampling + common grid), MultimodalFusionComposer (per-slice RGB rendering), TwoImageFusionViewer (amber/cyan edge fusion of two volumes) and ThreeImageOverlayViewer (fusion + semi-transparent red mask overlay).

Installation

# install from pypi
pip install ct_mri_dicom_nii_reader

# install from git repo
pip install -e .

Python >= 3.10 is required (the code uses modern type syntax and match).

Runtime dependencies (declared in pyproject.toml):

Dependency	Purpose
numpy	array storage and processing
simpleitk	DICOM and NIfTI decoding, resampling
pillow	GUI preview and display helpers
tkinter (optional)	preview windows; imported only when a preview is started

Quick start

from ct_mri_dicom_nii_reader import BodyDataLoaderManager

manager = BodyDataLoaderManager()

# 1. Load a CT volume from a DICOM series (any .dcm file inside the folder).
#    Voxels are resampled to 1 mm isotropic on an LPS grid.
ct = manager.load_file("path/to/dicom/IM0001.dcm")

# 2. Load an MRI volume from a NIfTI file.
mri = manager.load_file("path/to/image.nii.gz")

# 3. Save as the compact unified format, then reload it.
ct.save("out/ct_volume")               # writes out/ct_volume.ubd.npz
ct_again = manager.load_file("out/ct_volume.ubd.npz")

Working with BodyData

from ct_mri_dicom_nii_reader import BodyData, RoiRect

print(ct.get_type())        # 'ct' | 'mri' | 'mask'  (or None if uninitialized)
print(ct.get_size())        # (n_L, n_P, n_S)
print(ct.get_mmpd())        # mm per dot (voxel width)
print(ct.get_pos(0, 0, 0))  # value at voxel [l, p, s]

slice_l = ct.get_slice_l(100)      # BodyDataSlice view of a plane
slice_s = ct.get_slice_s(50)
slice_s.gui_preview()              # blocking 2-D preview window
ct.gui_preview()                   # blocking 3-D preview window with z slider

print(ct.percentile(1), ct.percentile(99))   # intensity percentiles
clone = ct.clone()                  # deep copy of the volume
mind = ct.get_mind()                # scalar MIND descriptor as BodyData

# Keep an x-z rectangle and blank everything else (air value is modality-aware).
roi = RoiRect(xmin=80, xmax=180, ymin=40, ymax=120)
part = ct.get_part(roi)

Core concepts

LPS axis order

All loaded volumes use the patient LPS convention:

• axis 0 -> Left, axis 1 -> Posterior, axis 2 -> Superior;
• volume[l, p, s] with increasing indices moving Left, Posterior, Superior;
• every output voxel is exactly mmpd mm on each side (isotropic).

DICOM rescaling (Rescale Slope / Intercept) is applied by SimpleITK/GDCM while reading, so CT values are Hounsfield units.

mmpd (mm per dot)

mmpd is the isotropic output voxel width in millimetres. All loaders resample the source image onto an mmpd-spaced LPS grid:

ct = manager.load_file("path/to/series.dcm", mmpd=1.5)  # 1.5 mm voxels

If omitted it defaults to 1.0 mm.

Image types

BodyData.get_type() returns one of:

Type	Meaning
"ct"	CT / DX / CR / CBCT volumes (HU)
"mri"	MR volumes
"mask"	segmentation labels or unknown modality

Masks whose values already lie in [0, 1] are automatically binarized to int8 (0 / 1) on load.

API reference

BodyDataLoaderManager

manager = BodyDataLoaderManager()
body_data = manager.load_file(filepath, mmpd=None)

Tries each registered loader (DICOM, NIfTI, unified) in turn and returns the first match. Raises NoAvailableDataLoader(filepath) when no loader can handle the path. mmpd defaults to 1.0.

Loader classes

BodyDataLoader is the abstract base class: subclasses implement check_match(filepath) and load_data(filepath). Users should call load_file(filepath) instead of load_data, which validates the match and post-processes masks.

Class	Handles
DicomBodyDataLoader	*.dcm (loads the folder containing the file)
NiiBodyDataLoader	*.nii, *.nii.gz
UnifiedBodyDataLoader	*.ubd.npz

All loaders accept get_mmpd() / set_mmpd() for the output voxel width.

BodyData

Method	Description
from_array(data, image_type, mmpd)	Wrap a NumPy array in LPS order
to_numpy(copy=False)	Return the internal 3-D NumPy array (copy only when copy=True)
get_initialized()	Whether image_type has been set
get_type()	"ct", "mri", "mask" or None
unify_to_mask()	Binarize values to int8 at threshold 0.5
get_size()	(n_L, n_P, n_S)
get_pos(l, p, s)	Value at a voxel
get_slice_l(l) / get_slice_p(p) / get_slice_s(s)	Plane views (BodyDataSlice)
get_mmpd() / set_mmpd(mmpd)	Voxel width accessors
percentile(idx)	idx in [0, 100] across all voxels
clone()	Deep copy
get_part(roi_rect)	Keep an x-z rectangle, fill elsewhere with the modality-specific air value
get_mind()	Scalar 3-D MIND descriptor of the volume
save(filepath)	Write .ubd.npz (extension added if missing, parent dirs created)
gui_preview()	Blocking 3-D slice browser window

BodyDataSlice

A 2-D plane view returned by BodyData.get_slice_*:

• get_pos(x, y), get_size() -> (x, y),
• to_numpy(copy=False) -> internal 2-D NumPy array,
• get_mmpd() / set_mmpd(), get_body_data() -> owning BodyData or None,
• clone() -> deep copy,
• gui_preview() -> blocking grayscale window.

RoiRect

roi = RoiRect(xmin, xmax, ymin, ymax)   # x-z rectangle in voxel indices

Accessors: get_xmin(), get_xmax(), get_ymin(), get_ymax().

Exceptions

Exception	Raised when
NoAvailableDataLoader(filepath)	No loader matches the file extension
DataLoaderNotMatch	A loader's load_file is called on a non-matching path
BodyDataNotInitialized	An operation requires an initialized BodyData
BodyDataTypeError(image_type)	The image type is not ct / mri / mask

Low-level loading functions

For full control, the implementation functions can be used directly (they are also what the loader classes call internally):

from ct_mri_dicom_nii_reader.body_data.body_data_imp.dicom_to_hu_lps import load_dicom_hu_lps
from ct_mri_dicom_nii_reader.body_data.body_data_imp.nifti_to_lps import load_nifti_lps

volume, metadata = load_dicom_hu_lps(
    "path/to/dicom_dir",
    mmpd=1.0,
    series_uid=None,            # optional DICOM Series Instance UID
    interpolation="linear",     # or "nearest"
    outside_hu=-1024.0,         # fill value outside the scanned field of view
    require_ct=True,            # reject non-CT series
    return_metadata=True,
)

volume, metadata = load_nifti_lps(
    "path/to/image.nii.gz",
    mmpd=1.0,
    interpolation="linear",     # use "nearest" for segmentation labels
    outside_value=None,         # defaults to -1024 (CT) or 0
)

DICOM metadata keys include series_uid, series_depth, series_voxel_count, modality, number_of_files, and the source geometry (source_size_xyz, source_spacing_xyz_mm, source_origin_lps_mm, source_direction_xyz_to_lps).

NIfTI metadata keys include modality ("CT", "MR" or "UNKNOWN"), modality_source and modality_confidence, the output geometry (origin_lps_mm, output_affine_lps_mm, outside_value) and the source geometry (source_size_xyz, source_spacing_xyz_mm, source_origin_lps_mm, source_direction_xyz_to_lps, source_affine_lps_mm, source_affine_ras_mm, json_sidecar).

DICOM series selection rules

When series_uid is not given, load_dicom_hu_lps:

1. lists every series in the directory,
2. excludes single-slice series,
3. picks the multi-slice series with the largest voxel count,
4. breaks ties deterministically by lexicographically smallest UID.

The loader class DicomBodyDataLoader relaxes require_ct to False and maps DICOM Modality to image type: CT/DX/CR/CBCT -> "ct", MR -> "mri", everything else -> "mask".

NIfTI modality detection order

1. same-name JSON sidecar (Modality or 00080060 key),
2. NIfTI header text (auxfile / descrip / intent ...),
3. filename (e.g. t1w, flair, ct),
4. intensity heuristic (P1 <= -500 and P99 >= 200 -> CT).

If nothing matches, the modality is "UNKNOWN" and the loader treats the image as a "mask".

Processing utilities

Function	Description
get_volume_percentile(volume, idx)	Percentile over all voxels of a 3-D array
compute_mind_image(volume, *, patch_radius, patch_sigma, neighbour_radius)	Scalar MIND descriptor in [0, 1], same shape as input
fill_outside_xz_rectangle(volume, l, r, t, b, air_value)	Keep [l, r) x [t, b) in the x-z plane, fill the rest

Import paths:

from ct_mri_dicom_nii_reader.body_data.body_data_imp.volume_percentile import get_volume_percentile
from ct_mri_dicom_nii_reader.body_data.body_data_imp.mind_3d import compute_mind_image
from ct_mri_dicom_nii_reader.body_data.body_data_imp.mask_xz_rectangle import fill_outside_xz_rectangle

GUI previews

from ct_mri_dicom_nii_reader.body_data.body_data_imp.slice_display import show_numpy_gray
from ct_mri_dicom_nii_reader.body_data.body_data_imp.numpy_3d_viewer import show_numpy_3d

show_numpy_gray(slice_2d, vmin, vmax)   # 2-D window, array[x, y] -> pixel (x, y)
show_numpy_3d(array_3d, vmin, vmax)     # 3-D window with z slider

Both are blocking (they open a Tk main loop) and use BodyData.gui_preview() / BodyDataSlice.gui_preview() as convenience wrappers.

The windows open at the default upscaled size produced by _upscale_for_display (smallest image side at least 512 px). Resizing the window rescales the image together with it while keeping its original aspect ratio.

Tkinter is optional and is imported lazily only when a preview starts; all data handling works headlessly without it. When tkinter is missing the preview functions raise an ImportError whose message includes the required install command:

pip install tk

(On Debian/Ubuntu, sudo apt-get install python3-tk works as an alternative.) To check availability from code without opening a window:

from ct_mri_dicom_nii_reader.body_data.body_data_imp._tk_gui import tkinter_available

print(tkinter_available())

Visualization

Three visualization modes are provided for BodyData volumes.

1. Single volume (grayscale)

BodyData.gui_preview() opens a grayscale window with a bottom slice slider — see show_numpy_3d above.

2. Two volumes (amber/cyan fusion)

TwoImageFusionViewer renders two volumes with the regknee-style amber/cyan edge fusion: the fixed volume contributes a dimmed grayscale background with amber edges, the moving volume contributes cyan edges, and where both edges overlap the colors add up to white.

from ct_mri_dicom_nii_reader import TwoImageFusionViewer

viewer = TwoImageFusionViewer(ct_body_data, mri_body_data)
viewer.gui_preview()   # blocking window with a bottom slice slider

If the two volumes differ in size or voxel spacing, they are first aligned at the LPS origin, resampled onto the finest unified spacing (the smallest mmpd), and placed into one common NumPy array whose per-axis lengths are the maxima of the resampled volumes. Regions covered by no volume hold the modality-specific air value (-1024 for CT, 0 for MRI and masks). The same combination is available programmatically:

from ct_mri_dicom_nii_reader import BodyDataCommonGrid, MultimodalFusionComposer

grid = BodyDataCommonGrid([ct_body_data, mri_body_data])
aligned = grid.get_aligned()          # list of BodyData on the common grid
grid.get_unified_mmpd()               # finest spacing
grid.get_common_size()                # (n_L, n_P, n_S) of the common array

composer = MultimodalFusionComposer(ct_body_data, mri_body_data)
rgb = composer.make_slice("s", 100)   # (height, width, 3) uint8 RGB slice

3. Three volumes (fusion + semi-transparent red mask)

ThreeImageOverlayViewer visualizes the first two volumes exactly as above; the third volume is aligned onto the same common grid and drawn as a red overlay on top of the fusion. The smallest mask value maps to a fully transparent overlay and the largest to 50% opacity; the blending happens in image (RGB) space.

from ct_mri_dicom_nii_reader import ThreeImageOverlayViewer

viewer = ThreeImageOverlayViewer(ct_body_data, mri_body_data, mask_body_data)
viewer.gui_preview()

Both viewer windows behave like the grayscale previews: they open at the default upscaled size, rescale with the window while preserving the aspect ratio, and raise an ImportError with the tkinter installation command when tkinter is missing.

Requirements

• Python >= 3.10
• numpy, SimpleITK, Pillow (see pyproject.toml)
• tkinter is optional and only needed for the preview GUIs; when it is missing the GUI functions report an error with the pip command to install it (pip install tk)

License

MIT. The classes were extracted from the regknee project (MIT, Copyright (c) 2026 GGN_2015) and keep their original design; see LICENSE.