Add phase correction options

[tsalo]

Summary

• Adds optional phase correction for complex-valued (magnitude + phase) diffusion data. After complex MP-PCA denoising (dwidenoise), the complex signal can be rephased and the real channel retained for downstream processing instead of the magnitude — avoiding the Rician noise-floor bias that magnitude data carries. Three rephasing methods are provided, the report gains a phase-correction panel, and the phase→radian conversion is aligned to the MRtrix convention. The code for these methods is adapted from fragrussu/MRItools.
• New CLI flag --dwi-phase-correction {none,tv,tvc,dc} (default none):
  • tv — total-variation rephasing on the magnitude/phase stack (Eichner 2015)
  • tvc — paper-faithful TV on the complex signal (Eichner 2015); wrap-free
  • dc — decorrelated-phase convolution filtering (Sprenger 2017); wrap-free
• When enabled, the real channel after rephasing replaces the magnitude downstream. The option requires --denoise-method dwidenoise and phase data (a BIDS part-phase file); the parser warns and ignores it otherwise. Tunables dwi_phase_tv_weight (default 6.0) and dwi_phase_dc_kernel (default Opt5) are exposed in config.

Changes by area

• New interface — qsiprep/interfaces/phase.py (515 lines)
  • rephase_tv, rephase_tv_complex, rephase_dc rephasing functions plus DC convolution kernels (B3–Opt5)
  • PhaseCorrect Nipype interface (method, tv_weight, dc_kernel) that writes the real channel and a report figure
• Workflow wiring — workflows/dwi/merge.py
  • init_dwi_denoising_wf branches: when phase correction is on, inserts a PhaseCorrect node (+ phasecorrection report sink) in place of ComplexToMagnitude; otherwise behavior is unchanged
  • Boilerplate text updated per method; warns that phase correction is skipped for concatenated (post-merge) series
• Derivatives — workflows/dwi/derivatives.py
  • Tags DWI/bval/bvec/gradient outputs with part-real when phase correction is active
• Phase scaling fix — interfaces/dwi_merge.py
  • PhaseToRad now rescales to [−π, π] (was [0, 2π]) to match MRtrix's complex dwidenoise convention (mrcalc … pi 4096 -div -mult -polar). This matters for the tv real channel, which is not invariant to a global phase offset.
• Gradient-file lookup — interfaces/images.py
  • New _find_gradient_file helper: complex acquisitions share one .bval/.bvec pair without the part- entity, so
• ConformDwi falls back to a part-stripped name.
• Report — data/reports-spec.yml
  • New "Complex Phase Correction" panel: real channel alongside the imaginary residual (low/high-b); a clean correction leaves only noise in the residual.
• Docs — docs/preprocessing.rst, docs/conf.py, boilerplate.bib
  • Reorganized "Denoising and Merging Images" into per-step prose sections (denoising, phase correction, Gibbs, bias correction, b0 harmonization, ordering, combining)
  • Fixed all Sphinx warnings: refreshed 7 .. workflow:: directives to current signatures (seeded qsiprep.config in conf.py), converted 2 non-buildable examples to static code blocks, repaired a cross-reference, and added jezzard1995/brett2001 bib entries (with eichner2015real, sprenger2017real, yeh2019differential citations)
• Tests (6 new files, ~375 lines)
  • test_interfaces_phase.py — each method recovers a known real signal and does not imprint magnitude anatomy onto the residual; DC kernels normalized; PhaseCorrect writes the real channel
  • test_interfaces_images.py — _find_gradient_file part-stripping fallback
  • test_parser_phase.py / test_config_phase.py — flag parsing, defaults, bad-value rejection
  • test_workflows_phase.py — node present when enabled, magnitude path unchanged when disabled, part-real tagging
  • test_boilerplate_phase.py — citations present
• Notes
  • Phase correction is not applied to concatenated DWI series (denoising runs after merge); magnitude is used and a warning is logged.
  • The Issue [BUG] Denoising visual report shows strange images when using complex MP-PCA #707 denoising-report fix (showing magnitude for complex inputs) is not in this branch — it lives on fix-complex-denoising-figure.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 78.53659% with 44 lines in your changes missing coverage. Please review.
✅ Project coverage is 47.92%. Comparing base (b142af8) to head (67735b6).

Files with missing lines	Patch %	Lines
qsiprep/interfaces/phase.py	79.22%	11 Missing and 5 partials ⚠️
qsiprep/utils/phase.py	81.17%	9 Missing and 7 partials ⚠️
qsiprep/cli/parser.py	16.66%	4 Missing and 1 partial ⚠️
qsiprep/interfaces/images.py	76.92%	2 Missing and 1 partial ⚠️
qsiprep/interfaces/dwi_merge.py	0.00%	2 Missing ⚠️
qsiprep/workflows/dwi/merge.py	84.61%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1054      +/-   ##
==========================================
+ Coverage   47.13%   47.92%   +0.78%     
==========================================
  Files          66       68       +2     
  Lines        9836    10035     +199     
  Branches     1084     1123      +39     
==========================================
+ Hits         4636     4809     +173     
- Misses       4979     4995      +16     
- Partials      221      231      +10     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[tsalo]

@fragrussu, this PR adapts your denoising code from imgrephaseTV.py and imgrephaseDC.py, and incorporates it into QSIPrep's denoising workflow. I was wondering, would you be willing to look at a few example outputs to see if they look right to you? I've noticed that there is more structure in the imaginary residuals for the low-b volumes than I'd expect, but this is the first time I've looked at outputs from this kind of method.

Also, I know that your code is licensed permissively, but would you like us to add you as a contributor to QSIPrep to reflect the fact that I'm incorporating your code with minimal modification?