LCS-Retrieval 0.1.0

LCS and MIST frameworks for single-exposure multi-contrast X-ray imaging: transmission, phase, isotropic and directional dark-field retrieval

🔬 LCS-Retrieval — Low Coherence System & MIST framework for multi-contrast X-ray imaging

---

LCS-Retrieval implements two complementary frameworks for simultaneous retrieval of transmission, phase, isotropic dark-field, and directional dark-field from X-ray images:

• LCS (Low Coherence System) — implicit sliding-window least-squares solver based on the TIE + Fokker–Planck model
• MIST (Multimodal Intrinsic Speckle Tracking) — phase and dark-field retrieval via local contrast statistics and FFT-based inversion

Both frameworks share a unified orientation analysis pipeline for the directional scattering tensor (theta, anisotropy, HSV visualization).

Compatible with speckle-based, modulation-based, and single-grid X-ray imaging setups.

Based on the paper:

Implicit Single-Exposure Retrieval of X-ray Phase, Absorption, and Anisotropic Dark-Field Samy KEFS, Chris Ninham, Clara Magnin, Sabine Rolland du Roscoat, Pierre Lhuissier, Emmanuel Brun IEEE Transactions on Medical Imaging, 2025

---

🚀 Overview

Conventional multi-contrast X-ray imaging requires multiple exposures or phase steps to separate absorption, phase, and scattering signals. LCS-Retrieval eliminates this need by reformulating the inverse problem locally within a sliding window around each pixel.

From a single pair of images (reference + sample), both methods retrieve:

Signal	LCS symbol	MIST symbol	Description
🔲 Transmission	T	Iob	Absorption contrast
🌊 Phase	dy, dx	phi	Phase / wavefront
🌀 Isotropic dark-field	DF	Deff	Overall scattering strength
🧭 Directional dark-field	Dxx, Dxy, Dyy	Dxx, Dxy, Dyy	Anisotropic scattering tensor

---

✳️ Key Features

• 🔹 Single-exposure retrieval — no phase stepping or mask motion required
• 🔹 LCS solver — sliding-window least-squares on the TIE + Fokker–Planck model
• 🔹 MIST solver — local contrast statistics + FFT-based Fokker–Planck inversion
• 🔹 Shared orientation pipeline — lcs.hsv_retrieval() and lcs.ddf_metrics() work on both LCS and MIST outputs
• 🔹 Multi-exposure mode supported for both methods (win_size=1)
• 🔹 External absorption injection — provide an independent T map for both LCS and MIST
• 🔹 HSV orientation visualization of the directional scattering tensor
• 🔹 Compatible with speckle-based, modulation-based, and single-grid setups

---

⚙️ Installation

▶️ From PyPI (recommended)

pip install LCS-Retrieval

To also install the dependencies for the example notebooks (matplotlib, fabio):

pip install LCS-Retrieval[examples]

▶️ From source

git clone https://github.com/MuguiwaraSamy/LCS-Retrieval.git
cd LCS-Retrieval
pip install -e .

Requirements: Python ≥ 3.9, NumPy ≥ 2.0, SciPy ≥ 1.10

---

🧠 Quick start — LCS

import lcs

# ref, sample: 2D arrays (H, W) or multi-exposure stacks (H, W, N)

# ── Isotropic dark-field ────────────────────────────────────────
result_df = lcs.lcs_df(ref, sample, win_size=5, alpha=1e-5)
# Output shape: (H', W', 4)
T  = result_df[..., 0]   # Transmission
dy = result_df[..., 1]   # Vertical refraction
dx = result_df[..., 2]   # Horizontal refraction
DF = result_df[..., 3]   # Isotropic dark-field

# ── Directional dark-field ──────────────────────────────────────
result_ddf = lcs.lcs_ddf(ref, sample, win_size=5, alpha=1e5)
# Output shape: (H', W', 6) → [T, dy, dx, Dxx, Dxy, Dyy]

# ── HSV orientation visualization ──────────────────────────────
rgb, metrics = lcs.hsv_retrieval(result_ddf, sigma=5.0)

---

🧠 Quick start — MIST

import mist
import lcs  # shared orientation pipeline

energy     = 8                    # keV
wavelength = 12.398e-10 / energy  # X-ray wavelength in metres
dist_det   = 0.8                  # metres

# ── Isotropic dark-field + phase ────────────────────────────────
result_df = mist.mist_df(ref, sample, dist_det=dist_det, wavelength=wavelength, win_size=3)
phi  = result_df['phi']   # Phase map
Deff = result_df['Deff']  # Isotropic dark-field

# ── Directional dark-field ──────────────────────────────────────
result_ddf = mist.mist_ddf(
    ref, sample, gamma_mat=100, dist_det=dist_det, wavelength=wavelength, win_size=3
)
# Output shape: (H', W', 6) → [thickness, Iob, G, Dxx, Dxy, Dyy]
thickness = result_ddf[..., 0]
Iob       = result_ddf[..., 1]
G         = result_ddf[..., 2]

# ── Shared orientation pipeline ─────────────────────────────────
rgb, metrics = lcs.hsv_retrieval(result_ddf, sigma=5.0)  # works on MIST too!

---

🗂️ Acquisition modes

Both methods support the same three modes:

Mode	Input shape	win_size
Single exposure + sliding window	(H, W)	odd integer ≥ 3
Multi-exposure (standard)	(H, W, N)	1
Multi-exposure + sliding window	(H, W, N)	odd integer ≥ 3

---

🧪 External absorption image

Both methods accept an optional absorption argument. The solver divides sample by absorption before retrieval, and returns the provided map (cropped to the valid window region) in the T/Iob channel.

# T_abs: independently measured absorption image, shape (H, W)

# LCS
result_df  = lcs.lcs_df( ref, sample, win_size=5, absorption=T_abs)
result_ddf = lcs.lcs_ddf(ref, sample, win_size=5, absorption=T_abs)

# MIST
result_df  = mist.mist_df( ref, sample, dist_det=dist_det, wavelength=wavelength, absorption=T_abs)
result_ddf = mist.mist_ddf(ref, sample, gamma_mat=100, dist_det=dist_det,
                            wavelength=wavelength, absorption=T_abs)

---

🔗 Shared orientation pipeline

Both lcs.lcs_ddf and mist.mist_ddf return a (H', W', 6) stack whose last 3 channels are always the scattering tensor [Dxx, Dxy, Dyy]. This means the orientation analysis functions work identically on both:

# Works with both lcs.lcs_ddf() and mist.mist_ddf() outputs
rgb, metrics = lcs.hsv_retrieval(result_ddf, sigma=5.0)
metrics      = lcs.ddf_metrics(result_ddf, sigma=5.0)

# Or directly on tensor components
metrics = lcs.tensor_metrics(Dxx, Dxy, Dyy, sigma=5.0)

metrics contains: theta, theta_smooth, coherence, aniso, traceS, major, minor, lambda_max, lambda_min, intensity

---

📖 API reference

LCS

lcs.lcs_df(ref, sample, *, win_size=5, alpha=1e-5, absorption=None)

Parameter	Type	Default	Description
ref	ndarray (H,W) or (H,W,N)	—	Reference image(s)
sample	ndarray (H,W) or (H,W,N)	—	Sample image(s)
win_size	int	5	Sliding window size (odd)
alpha	float	1e-5	Tikhonov regularization
absorption	ndarray (H,W) or None	None	External absorption image

Returns ndarray (H', W', 4) → [T, dy, dx, DF]

lcs.lcs_ddf(ref, sample, *, win_size=5, alpha=1e5, absorption=None)

Returns ndarray (H', W', 6) → [T, dy, dx, Dxx, Dxy, Dyy]

---

MIST

mist.mist_df(ref, sample, dist_det, wavelength, *, win_size=3, sig_scale=0.0, median_filter_size=0, alpha=1e5, absorption=None)

Parameter	Type	Default	Description
ref	ndarray (H,W) or (H,W,N)	—	Reference image(s)
sample	ndarray (H,W) or (H,W,N)	—	Sample image(s)
dist_det	float	—	Sample-to-detector distance (m)
wavelength	float	—	X-ray wavelength in metres (12.398e-10 / energy_keV)
win_size	int	3	Sliding window size (odd)
sig_scale	float	0.0	Gaussian high-pass filter scale
median_filter_size	int	0	Median filter on Deff (0 = off)
alpha	float	1e5	Tikhonov regularization
absorption	ndarray (H,W) or None	None	External absorption image

Returns dict → {'phi', 'Deff', 'phi_laplacian'}

mist.mist_ddf(ref, sample, gamma_mat, dist_det, wavelength, *, win_size=3, sig_scale=0.0, alpha=1e5, absorption=None)

Parameter	Type	Default	Description
gamma_mat	float	—	Material delta/beta ratio
dist_det	float	—	Sample-to-detector distance (m)
wavelength	float	—	X-ray wavelength in metres (12.398e-10 / energy_keV)
sig_scale	float	0.0	Gaussian high-pass filter scale
absorption	ndarray (H,W) or None	None	External absorption image

Returns ndarray (H', W', 6) → [thickness, Iob, G, Dxx, Dxy, Dyy]

---

Shared orientation functions

lcs.hsv_retrieval(result_ddf, *, sigma=5.0, percentile=99.0, sat_min=0.0, sat_max=1.0, val_min=0.0, val_max=1.0)

Returns (rgb, metrics) — works on both LCS and MIST outputs.

lcs.ddf_metrics(result_ddf, *, sigma=5.0)

Returns metrics dict — works on both LCS and MIST outputs.

lcs.tensor_metrics(Dxx, Dxy, Dyy, *, sigma=5.0)

Returns metrics dict directly from tensor components.

---

📘 Example notebooks

Notebook	Description
examples/example_lcs.ipynb	LCS isotropic & directional dark-field
examples/example_mist.ipynb	MIST phase, isotropic & directional dark-field

---

🖥️ ImageJ Plugin

A dedicated ImageJ/Fiji plugin for LCS retrieval is also available: 👉 github.com/MuguiwaraSamy/LCS-ImageJ

---

📄 Citation

If you use this package in your work, please cite:

@article{kefs2025lcs,
  title   = {Implicit Single-Exposure Retrieval of {X}-ray Phase, Absorption, and Anisotropic Dark-Field},
  author  = {Kefs, Samy and Ninham, Chris and Magnin, Clara and
             Rolland du Roscoat, Sabine and Lhuissier, Pierre and Brun, Emmanuel},
  journal = {IEEE Transactions on Medical Imaging},
  year    = {2025},
}

---

📜 License

Distributed under the MIT License. © 2025 Samy KEFS. All rights reserved.