pf_MIDAS.py User Manual

Version: 11.0 Contact: hsharma@anl.gov

Note

For standard (box-beam) FF-HEDM analysis, see FF_Analysis.md. PF-HEDM (Point-Focus / Pencil-beam / Scanning FF-HEDM) processes data from multiple sample positions to reconstruct a spatially-resolved microstructure map.

1. Introduction

pf_MIDAS.py is the driver script for Point-Focus (Scanning) FF-HEDM analysis within MIDAS. Unlike standard box-beam FF-HEDM, scanning FF-HEDM translates the sample across a focused beam, collecting diffraction data at each position. This allows voxel-level orientation mapping similar to EBSD but using high-energy X-rays for non-destructive 3D characterization.

The script:

Runs peak search on each scan position in parallel.
Combines spot data across all positions.
Performs scanning-mode indexing and refinement.
Optionally reconstructs tomographic sinograms (inverse Radon transform) for each grain.
Produces a microstrFull.csv and microstructure.hdf with the spatially-resolved microstructure.

2. Prerequisites

A working MIDAS installation.
Raw diffraction data (GE, HDF5, or pre-built Zarr-ZIP).
A calibrant-derived parameter file (see FF_Calibration.md).
A positions.csv file listing Y-positions (one per line, negative w.r.t. motor position).
Parameter file and positions.csv must be in the same directory.
Python environment with: parsl, numpy, pandas, scikit-image, Pillow, h5py, zarr, numba.

3. Command-Line Arguments

python pf_MIDAS.py -paramFile <param.txt> [options]

Argument	Type	Default	Description
`-paramFile`	`str`	Required	Parameter file name (basename only, not full path).
`-resultDir`	`str`	`''` (cwd)	Output directory for results.
`-nCPUs`	`int`	`32`	Number of CPUs per node for peak search and indexing.
`-nCPUsLocal`	`int`	`4`	Number of local CPUs for non-parallelized tasks.
`-nNodes`	`int`	`1`	Number of compute nodes.
`-machineName`	`str`	`local`	Execution target: `local`, `orthrosnew`, `orthrosall`, `umich`, `marquette`.
`-doPeakSearch`	`int`	`1`	`1` = run peak search; `0` = skip (already done); `-1` = re-process peak output without re-running search.
`-convertFiles`	`int`	`1`	`1` = convert raw to Zarr-ZIP; `0` = use existing.
`-runIndexing`	`int`	`1`	`1` = run indexing; `0` = skip indexing.
`-oneSolPerVox`	`int`	`1`	`1` = single orientation per voxel; `0` = allow multiple orientations per voxel.
`-doTomo`	`int`	`1`	`1` = reconstruct tomographic sinograms; `0` = skip. Only for `-oneSolPerVox 1`.
`-normalizeIntensities`	`int`	`2`	`0` = equivalent grain size; `1` = normalize by powder intensity; `2` = integrated intensity.
`-numFrameChunks`	`int`	`-1`	Chunk data for low-RAM systems. `-1` disables.
`-preProcThresh`	`int`	`-1`	Pre-processing threshold above dark. `-1` disables.
`-startScanNr`	`int`	`1`	First scan number to process (for partial peak search).
`-minThresh`	`int`	`-1`	Filter peaks with maxIntensity below this value. `-1` disables.
`-micFN`	`str`	`''`	Path to a `.mic` file for guided indexing.
`-grainsFN`	`str`	`''`	Path to a grains file for seed-based indexing.
`-omegaFile`	`str`	`''`	Override omega values (one per scan, text file).
`-sinoType`	`str`	`raw`	Sinogram variant to use for reconstruction. Valid: `raw`, `norm`, `abs`, `normabs`. See Section 7.
`-sinoSource`	`str`	`tolerance`	Sinogram spot source: `tolerance` = match all spots by angular tolerance (default); `indexing` = use only spots from per-voxel indexing results (cleaner).
`-reconMethod`	`str`	`fbp`	Reconstruction algorithm: `fbp` (Filtered Back-Projection via gridrec), `mlem` (Maximum Likelihood EM), `osem` (Ordered Subsets EM). See Section 8.3.
`-mlemIter`	`int`	`50`	Number of MLEM/OSEM iterations (only used when `-reconMethod` is `mlem` or `osem`).
`-osemSubsets`	`int`	`4`	Number of ordered subsets for OSEM (only used when `-reconMethod` is `osem`).
`-resume`	`str`	`''`	Path to a pipeline H5 to resume from. Auto-detects the last completed stage.
`-restartFrom`	`str`	`''`	Explicit stage to restart from. Valid stages: `hkl`, `peak_search`, `merge`, `params_rewrite`, `indexing`, `refinement`, `find_multiple_solutions`, `consolidation`.

The following flags appear only when the optional sr-midas pip package is installed (see Section 3a):

Argument	Type	Default	Description
`-runSR`	`int`	`0`	`1` = replace per-scan MIDAS peak search with the sr-midas super-resolution pipeline. Requires `-doPeakSearch 0`.
`-srfac`	`int`	`8`	Super-resolution upscale factor. Choices: `2`, `4`, `8`.
`-SRconfig_path`	`str`	`auto`	Path to a custom sr-midas config JSON. `auto` uses the config bundled with sr-midas.
`-saveSRpatches`	`int`	`0`	`1` = save the predicted super-resolved patches to disk (`SR_out/SR_patches/` per scan).
`-saveFrameGoodCoords`	`int`	`0`	`1` = save per-frame `goodCoords` maps per scan.

Example

# Full analysis with tomographic reconstruction:
python pf_MIDAS.py -paramFile ps_pf.txt -resultDir ~/results/ -nCPUs 16 -doTomo 1

# Skip peak search (already done):
python pf_MIDAS.py -paramFile ps_pf.txt -doPeakSearch 0 -nCPUs 32

# Multiple orientations per voxel (no tomography):
python pf_MIDAS.py -paramFile ps_pf.txt -oneSolPerVox 0 -doTomo 0

# Resume from the last completed stage:
python pf_MIDAS.py -paramFile ps_pf.txt -resume /path/to/pipeline.h5

# Restart from indexing:
python pf_MIDAS.py -paramFile ps_pf.txt -restartFrom indexing

# Use MLEM reconstruction instead of FBP:
python pf_MIDAS.py -paramFile ps_pf.txt -doTomo 1 -reconMethod mlem -mlemIter 80

# Use OSEM (faster convergence) with normalized sinograms:
python pf_MIDAS.py -paramFile ps_pf.txt -doTomo 1 -reconMethod osem -osemSubsets 4 -sinoType norm

3a. Super-Resolution Peak Search (optional, `sr-midas`, experimental for PF)

pf_MIDAS.py can delegate the peak-fitting step inside each scan's parallel_peaks Parsl task to sr-midas, an optional PyPI package that upscales detector patches with a cascaded CNN before fitting peaks. The surrounding per-scan stages (zip generation, GetHKLListZarr, MergeOverlappingPeaksAllZarr, omega handling, CalcRadiusAllZarr, FitSetupZarr, intensity normalization, consolidated-CSV write) still run unchanged, so downstream indexing / refinement / tomography are unaffected.

Warning

PF-HEDM integration is EXPERIMENTAL. The sr-midas README only documents FF-HEDM; the PF per-scan integration was inferred from the on-disk layout and has not been validated by the sr-midas authors. Verify outputs against a standard peak-search reference run before relying on them scientifically. A loud per-layer warning is printed to the log each time SR runs.

Install (one-time, in a Python 3.12.4 environment, with PyTorch that matches your GPU):

pip install sr-midas

When the package is importable, the five -runSR* flags above are automatically registered and a banner prints at startup:

SR-MIDAS: available (version 0.1.1).

If sr-midas isn't installed the flags don't appear in --help and pf_MIDAS.py behaves exactly as before.

GPU strongly recommended. sr-midas auto-detects CUDA; on a CPU-only host the driver logs a prominent warning and inference is 10–100× slower. Because PF runs one SR pass per scan (potentially dozens to hundreds), the CPU penalty compounds quickly.

Usage:

# Replace per-scan peak fitting with sr-midas (srfac=8 default):
python pf_MIDAS.py -paramFile ps_pf.txt -doPeakSearch 0 -runSR 1

# Lower upscale factor for faster per-scan inference:
python pf_MIDAS.py -paramFile ps_pf.txt -doPeakSearch 0 -runSR 1 -srfac 4

# Save diagnostic SR patches and goodCoords maps:
python pf_MIDAS.py -paramFile ps_pf.txt -doPeakSearch 0 -runSR 1 \
    -saveSRpatches 1 -saveFrameGoodCoords 1

Guardrails:

-runSR 1 with -doPeakSearch 1 exits with an error — sr-midas replaces the per-scan peak-fit step.
-runSR 1 without sr-midas installed exits with a pip install sr-midas hint.
Each scan's .MIDAS.zip must already exist (either from a prior run or from this invocation's zip generation). If the zarr is missing for a scan, that layer's SR task fails fast with a clear error.

4. Parameter File Reference

The parameter file uses the same format as FF_Analysis.md. The following additional keys are specific to PF-HEDM:

Key	Type	Description
`nScans`	`int`	Number of scan positions (must match lines in `positions.csv`)
`BeamSize`	`float`	Beam size (μm). Multiplied by `nStepsToMerge` if merging is used.
`nStepsToMerge`	`int`	Number of adjacent scans to merge (0 = no merging)
`MaxAng`	`float`	Maximum misorientation angle for grain matching (degrees)
`TolOme`	`float`	Omega tolerance for spot matching (degrees)
`TolEta`	`float`	Eta tolerance for spot matching (degrees)
`OverAllRingToIndex`	`int`	Primary ring for indexing
`MicFile`	`str`	Path to `.mic` file for guided indexing (optional)
`GrainsFile`	`str`	Path to seed grains file (optional)
`MinMatchesToAcceptFrac`	`float`	Minimum completeness fraction to accept a solution (multi-solution mode)

All core FF-HEDM parameters (Lsd, Wavelength, BC, tilts, LatticeParameter, SpaceGroup, etc.) apply identically — see FF_Analysis.md § 4.

5. Input Files

positions.csv

One Y-position per line (μm), negative with respect to the motor position. The number of lines must equal nScans.

-50.0
-40.0
-30.0
...

Parameter File

Must be in the same directory as positions.csv. The parameter file name passed via -paramFile should be the basename only (not a full path).

6. Workflow Architecture

flowchart TD
    A[Parameter File + positions.csv] --> B[Per-position peak search<br/>parallel_peaks × nScans]
    B --> C{Merge scans?}
    C -->|Yes| D[mergeScansScanning<br/>nStepsToMerge]
    C -->|No| E[Combined spot data]
    D --> E
    E --> F[SaveBinDataScanning<br/>Bin combined data]
    F --> G[IndexerScanningOMP<br/>Find orientations]
    G --> H{One sol per voxel?}
    H -->|Yes| I[findSingleSolutionPF<br/>Select best orientation]
    H -->|No| J[findMultipleSolutionsPF<br/>Allow multiple orientations]
    I --> K{Do tomography?}
    K -->|Yes| L[Sinogram variants<br/>raw / norm / abs / normabs]
    K -->|No| M[FitOrStrainsScanningOMP<br/>Refine positions & strains]
    L --> L2{Recon method?}
    L2 -->|FBP| L3[gridrec via MIDAS_TOMO]
    L2 -->|MLEM| L4[mlem_recon.mlem]
    L2 -->|OSEM| L5[mlem_recon.osem]
    L3 --> N[microstructure.hdf + microstrFull.csv]
    L4 --> N
    L5 --> N
    J --> M
    L3 --> M
    L4 --> M
    L5 --> M
    M --> N

    style A fill:#1a1a2e,stroke:#e94560,color:#fff
    style N fill:#1a1a2e,stroke:#00d4aa,color:#fff

Stage Descriptions

Stage	Binary / Function	Description
Peak Search	`parallel_peaks()` (Parsl)	Runs per-position pipeline: ZIP generation, HKL list, peak search, merge, radius, fit setup
Scan Merging	`mergeScansScanning`	Merges adjacent scan positions to increase signal (optional)
Binning	`SaveBinDataScanning`	Bins spots from all scan positions for efficient search
Indexing	`IndexerScanningOMP`	Scanning-mode indexing with position awareness
Single Solution	`findSingleSolutionPF`	Selects best unique orientation per voxel
Multi Solution	`findMultipleSolutionsPF`	Allows overlapping grains in a single voxel
Tomography	FBP / MLEM / OSEM	Reconstructs tomographic images per grain from sinograms (see Section 8.3)
Refinement	`FitOrStrainsScanningOMP`	Refines orientations, positions, and strains

Pipeline Stage Order (for `-restartFrom`)

The internal stage order used by -restartFrom and -resume is:

hkl
peak_search
merge
params_rewrite
indexing
refinement
find_multiple_solutions
consolidation

Restarting from a given stage re-runs that stage and all subsequent stages.

7. Output Files

Directory Structure

<resultDir>/
├── positions.csv                  # Input positions
├── paramstest.txt                 # Auto-generated parameter file
├── hkls.csv                       # HKL reflections
├── <startNr>/                     # Per-scan-position subdirectories
│   ├── paramstest.txt
│   ├── <filestem>_NNNNNN.MIDAS.zip
│   └── output/
├── InputAllExtraInfoFittingAll0.csv  # Combined spots (scan 0)
├── InputAllExtraInfoFittingAll1.csv  # Combined spots (scan 1)
├── ...
├── SpotsToIndex.csv               # Spots selected for indexing
├── UniqueOrientations.csv         # Unique grain orientations
├── Output/
│   ├── UniqueIndexSingleKey.bin   # Single-solution voxel map
│   └── IndexBest_voxNr_*.bin      # Best index per voxel
├── Results/
│   └── *.csv                      # Per-voxel refinement results
├── Sinos/                         # Tomographic sinograms (if -doTomo 1)
│   ├── sino_raw_grNr_NNNN.tif    # Raw intensity sinogram
│   ├── sino_norm_grNr_NNNN.tif   # Normalized (I/Imax) sinogram
│   ├── sino_abs_grNr_NNNN.tif    # Absorption (exp(-I)) sinogram
│   └── sino_normabs_grNr_NNNN.tif # Normalized absorption (exp(-I/Imax)) sinogram
├── Thetas/                        # Per-grain theta arrays
│   └── thetas_grNr_*.txt
├── Recons/                        # Reconstructed images
│   ├── recon_grNr_*.tif           # Per-grain reconstruction
│   ├── Full_recon_max_project.tif
│   ├── Full_recon_max_project_grID.tif
│   ├── all_recons_together.tif
│   ├── microstrFull.csv           # ★ Final microstructure CSV
│   └── microstructure.hdf         # ★ Final HDF5 output
├── output/
│   ├── mapping_out.csv / mapping_err.csv
│   ├── indexing_out*.csv / indexing_err*.csv
│   └── refining_out*.csv / refining_err*.csv
└── processing.log

Sinogram Variants

When -doTomo 1 is enabled, the workflow saves four sinogram variants per grain as TIF files in the Sinos/ subdirectory. These are produced by findSingleSolutionPFRefactored and written out by save_sinogram_variants():

Variant	Filename Pattern	Description
`raw`	`sino_raw_grNr_NNNN.tif`	Raw intensity values from matched spots
`norm`	`sino_norm_grNr_NNNN.tif`	Normalized intensity (I / I_max per grain)
`abs`	`sino_abs_grNr_NNNN.tif`	Absorption-like transform: exp(-I)
`normabs`	`sino_normabs_grNr_NNNN.tif`	Normalized absorption: exp(-I / I_max)

Each sinogram TIF has shape (nScans, nSpots) and stores double-precision values. The -sinoType argument controls which variant is fed into the reconstruction step.

microstrFull.csv Column Format (43 columns)

The output CSV contains one row per indexed voxel with the following columns:

Columns	Name	Description
1	`SpotID`	Spot/Voxel identifier
2–10	`O11`–`O33`	Orientation matrix (3×3, row-major)
11	`SpotID`	(repeated)
12–14	`x`, `y`, `z`	Position (μm)
15	`SpotID`	(repeated)
16–21	`a`, `b`, `c`, `alpha`, `beta`, `gamma`	Fitted lattice parameters
22	`SpotID`	(repeated)
23	`PosErr`	Position error
24	`OmeErr`	Omega error
25	`InternalAngle`	Internal angle metric
26	`Radius`	Equivalent grain radius
27	`Completeness`	Fraction of expected spots matched
28–36	`E11`–`E33`	Strain tensor (6 unique components)
37–39	`Eul1`, `Eul2`, `Eul3`	Euler angles (Bunge convention, degrees)
40–43	`Quat1`–`Quat4`	Quaternion (fundamental region)

microstructure.hdf

HDF5 file with two datasets:

microstr: Full results array (same as microstrFull.csv)
images: 3D array (23, nScans, nScans) with spatially-resolved data suitable for imaging — includes ID, quaternion, position, lattice parameters, strain, completeness.

8. Technical Implementation Details

8.1. Spatially-Aware Indexing (`IndexerScanningOMP`)

Unlike standard box-beam indexing, the scanning indexer accounts for the sample's translation across the beam.

Dynamic Geometry: For every candidate voxel at position $(x, y)$, the diffraction spot projection is recalculated. The expected detector $Y$ position ($Y_{det}$) is modified by the sample translation projected onto the detector plane:

$$Y_{det} = Y_{theor} + \frac{x \cdot \sin \omega + y \cdot \cos \omega}{px_{size}}$$

Voxel Grid: The software discretizes the sample space into a grid defined by the BeamSize. It systematically tests orientations at each grid point, effectively performing a "diffraction-based raster scan."

8.2. Sinogram Generation

When -doTomo 1 is enabled, sinograms are constructed before reconstruction:

Per-grain aggregation: For each identified grain, the script aggregates intensity metrics across all scan positions and rotation angles ($\omega$). This forms a sinogram where the vertical axis is the scan position and the horizontal axis is the projection angle.
Four variants: The sinogram is saved in four processing combinations (raw, normalized, absorption, normalized-absorption) to allow flexibility in the downstream reconstruction. See Sinogram Variants in Section 7.
Variant selection: The -sinoType argument controls which variant is used for reconstruction. For samples with strong absorption contrast, abs or normabs may produce better grain shape reconstructions. For most diffraction-intensity cases, raw (default) or norm work well.

8.3. Reconstruction Methods (FBP, MLEM, OSEM)

The workflow supports three tomographic reconstruction algorithms, selected via -reconMethod:

FBP (Filtered Back-Projection) — default

The standard analytical reconstruction method. Uses the gridrec algorithm via MIDAS_TOMO (midas_tomo_python.run_tomo_from_sinos). FBP is fast and produces good results when the sinogram has dense, uniform angular coverage.

python pf_MIDAS.py -paramFile ps_pf.txt -doTomo 1 -reconMethod fbp

MLEM (Maximum Likelihood Expectation Maximization)

An iterative statistical reconstruction method (utils/mlem_recon.mlem) that is better suited for sparse or irregularly sampled sinograms. MLEM naturally handles:

Missing projections: Only sinogram rows with non-zero data are used; missing angles do not introduce artifacts.
Non-uniform angular sampling: No assumption of evenly spaced projection angles.
Positivity constraint: The multiplicative update inherently produces non-negative reconstructions.
Poisson noise model: More appropriate than FBP's implicit Gaussian assumption for photon-counting data.

Each iteration forward-projects the current estimate, computes the ratio of measured to projected data, back-projects the ratio, and applies a multiplicative correction. Convergence is controlled by -mlemIter (default: 50).

python pf_MIDAS.py -paramFile ps_pf.txt -doTomo 1 -reconMethod mlem -mlemIter 80

OSEM (Ordered Subsets Expectation Maximization)

An accelerated variant of MLEM (utils/mlem_recon.osem) that divides the projection angles into interleaved subsets and updates the estimate after each subset rather than after a full pass. This converges approximately n_subsets times faster than standard MLEM, making it practical for large datasets. Controlled by -mlemIter (iterations over all subsets) and -osemSubsets (number of subsets, default: 4).

python pf_MIDAS.py -paramFile ps_pf.txt -doTomo 1 -reconMethod osem -mlemIter 50 -osemSubsets 4

When to use each method

Scenario	Recommended Method
Dense angular coverage, fast results needed	`fbp`
Sparse/irregular omega angles, missing projections	`mlem` or `osem`
Large number of grains, need speed	`osem` (faster convergence than `mlem`)
Best quality from limited data, time not critical	`mlem` with high iteration count

9. Troubleshooting

Issue	Likely Cause	Resolution
`positions.csv` not found	Wrong directory or missing file	Ensure `-paramFile` and `positions.csv` are in the same folder
`Failed at generateZip for layer N`	Raw data not found	Check `RawFolder`, `FileStem`, `StartFileNrFirstLayer`
Empty `InputAllExtraInfoFittingAll*.csv`	No peaks found	Lower `RingThresh`, check detector geometry
No sino files found	`findSingleSolutionPF` found no orientations	Lower `MinNrSpots`, check `MaxAng`
Out of memory	Large number of scans × rings	Use `-numFrameChunks` and fewer `-nCPUs`
`Error reading sino data`	Mismatched `nScans` vs. `positions.csv`	Verify `nScans` matches number of positions
`Invalid restart stage`	Wrong stage name in `-restartFrom`	Use one of: `hkl`, `peak_search`, `merge`, `params_rewrite`, `indexing`, `refinement`, `find_multiple_solutions`, `consolidation`

10. GPU Acceleration & Consolidated I/O

Consolidated Binary I/O

Per-voxel file I/O has been replaced with consolidated binary format:

3 binary files per scan: IndexBest_all.bin, IndexKey_all.bin, IndexBest_IDs_all.bin
Reduces filesystem overhead from ~30K+ small files to 3 files
Uses IndexerConsolidatedIO.h with VoxelAccumulator (writer) and ConsolidatedReader (mmap-based reader) for O(1) voxel access
pf_MIDAS.py reads consolidated IndexBest_all.bin via numpy for .mic file generation

GPU Acceleration

Enable GPU-accelerated scanning indexer and fitter:

python pf_MIDAS.py -paramFN params.txt -useGPU 1

IndexerScanningGPU supports three modes: spot-driven with beam proximity filter, MicFile-seeded, and GrainsFile-seeded. FitOrStrainsScanningGPU reads consolidated indexer output.

Resume/Restart

--resume and --restartFrom flags are fully wired with _should_run() gates for proper stage-skipping.

See GPU_Acceleration.md for full GPU documentation.

11. See Also

FF_Analysis.md — Standard (box-beam) FF-HEDM analysis
FF_Calibration.md — Geometry calibration
PF_Interactive_Plotting.md — Interactive sinogram, intensity & tomo viewer for PF-HEDM
FF_Interactive_Plotting.md — Visualizing FF-HEDM results
Tomography_Reconstruction.md — MIDAS tomography reconstruction
Forward_Simulation.md — Forward simulation for validation
NF_Analysis.md — Near-field HEDM reconstruction
README.md — High-level MIDAS overview and manual index

If you encounter any issues or have questions, please open an issue on this repository.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

pf_MIDAS.py User Manual

1. Introduction

2. Prerequisites

3. Command-Line Arguments

Example

3a. Super-Resolution Peak Search (optional, `sr-midas`, experimental for PF)

4. Parameter File Reference

5. Input Files

positions.csv

Parameter File

6. Workflow Architecture

Stage Descriptions

Pipeline Stage Order (for `-restartFrom`)

7. Output Files

Directory Structure

Sinogram Variants

microstrFull.csv Column Format (43 columns)

microstructure.hdf

8. Technical Implementation Details

8.1. Spatially-Aware Indexing (`IndexerScanningOMP`)

8.2. Sinogram Generation

8.3. Reconstruction Methods (FBP, MLEM, OSEM)

FBP (Filtered Back-Projection) — default

MLEM (Maximum Likelihood Expectation Maximization)

OSEM (Ordered Subsets Expectation Maximization)

When to use each method

9. Troubleshooting

10. GPU Acceleration & Consolidated I/O

Consolidated Binary I/O

GPU Acceleration

Resume/Restart

11. See Also

FilesExpand file tree

PF_Analysis.md

Latest commit

History

PF_Analysis.md

File metadata and controls

pf_MIDAS.py User Manual

1. Introduction

2. Prerequisites

3. Command-Line Arguments

Example

3a. Super-Resolution Peak Search (optional, sr-midas, experimental for PF)

4. Parameter File Reference

5. Input Files

positions.csv

Parameter File

6. Workflow Architecture

Stage Descriptions

Pipeline Stage Order (for -restartFrom)

7. Output Files

Directory Structure

Sinogram Variants

microstrFull.csv Column Format (43 columns)

microstructure.hdf

8. Technical Implementation Details

8.1. Spatially-Aware Indexing (IndexerScanningOMP)

8.2. Sinogram Generation

8.3. Reconstruction Methods (FBP, MLEM, OSEM)

FBP (Filtered Back-Projection) — default

MLEM (Maximum Likelihood Expectation Maximization)

OSEM (Ordered Subsets Expectation Maximization)

When to use each method

9. Troubleshooting

10. GPU Acceleration & Consolidated I/O

Consolidated Binary I/O

GPU Acceleration

Resume/Restart

11. See Also

3a. Super-Resolution Peak Search (optional, `sr-midas`, experimental for PF)

Pipeline Stage Order (for `-restartFrom`)

8.1. Spatially-Aware Indexing (`IndexerScanningOMP`)