DICOM extract: multi-echo plane detection, modality fallback, and 9 smaller fixes by robert-graf · Pull Request #152 · Hendrik-code/TPTBox

robert-graf · 2026-09-15T08:24:03Z

Summary

Cleanup pass on TPTBox/core/dicom/. Splits into 10 focused commits (plus the pre-existing 07aac40 fix multi echo bug). Every commit is reviewable on its own.

Correctness fixes

88ab7e8 — _extract_nii_from_dicom returns False on unknown exceptions, deletes the half-written sidecar, and logs. Previously the bare except Exception fell through to return True, so _add_grid_info_to_json / _split_multi_echo_dixon then ran on a NIfTI that was never written.
391a28e — _classic_get_grouped_dicoms no longer produces NaN directions on multi-echo series. New helper _split_by_echo_numbers runs stack detection on a single-echo subset, then re-attaches remaining echoes by ImagePositionPatient.
f0c70f4 — _split_multi_echo_dixon triggers on any 4-D output (not just part in ("magnitude", "phase")). Also switched frame/TE iteration to zip_strict.
0894e46 — _find_all_files yields Path instances one at a time instead of yielding the raw list[Path] first, which downstream code cast to str and broke on.
7f9e38b — _inc_key caps its retry loop at 10 000 iterations and raises RuntimeError with the offending key instead of spinning forever on a broken path_exists callback.
a09de6a — get_plane_dicom's real call now uses 0.8 (matches the docstring default); the bare-except now logs a warning so downstream acq-None files are traceable.

Perf / usability

4c63431 — _read_dicom_files gates on a _looks_like_dicom check (extension blocklist + DICM magic at offset 128) before pydicom.dcmread. Skips reports, thumbnails, JSON sidecars, etc. on mixed source trees.
7a7ade2 — Docs: n_cpu behaviour spelled out (1 sequential / >1 N threads / None → ThreadPoolExecutor-Default ≈ min(32, cpu_count + 4)).

Features

e1cddbc — Expanded map_series_description_to_file_format_default:
- Quantitative maps (MP2RAGE, T2*, R2*, T1map, T2map, T2starmap, SWI) placed before the greedy .*t2.*/.*t1.* catch-alls that were misclassifying them.
- Multi-echo VIBE variants (ME_vibe_fatquant_*, mevibe, fatquant, fatfrac, pdff, GE ideal).
- Philips localizers (survey, refscan, smartexam).
5280eb8 — Modality fallback + BIDS format whitelist + view/laterality/bodypart:
- bids_constants.formats grows by xray, us, nm, sc, photo, endoscopy, rtimage, rtstruct, rtdose, rtplan, ot.
- extract_keys_from_json no longer raises NotImplementedError for CR/DX/RG/PX/IO/MG/US/NM/SC/OP/XC/ES/RT*/OT — sensible defaults per DICOM modality group, with a warning + tag-based fallback for unknowns.
- Two helpers _apply_view_keys and _apply_bodypart_key lift ViewPosition, Laterality/ImageLaterality, and BodyPartExamined into acq / part / desc, so MG four-view screenings, bilateral radiographs, ophthalmic photos, and multi-region NM/US sessions no longer collapse onto identical filenames.

Test plan

pytest unit_tests/ — 595 passed, 5 skipped (unchanged from base).
End-to-end run through the GLP1/RESIST DICOM tree, confirming the 6-echo Philips series is now labelled acq-ax (not acq-iso) and the classic stack detector no longer over-splits it.

🤖 Generated with Claude Code

…-Reg

Roll the internal-scripts changes together with the new precompute_grid_info_parallel helper in _load_nako_wh, which fans per-file _add_grid_info_to_json calls out to a ProcessPoolExecutor so the sidecar grid cache is populated up front.

…-Reg

The bare `except Exception: print(nii_path)` at the end of _extract_nii_from_dicom still fell through to `return True`. Any unknown conversion error (network, out of disk, dicom2nifti internal bug) was silently treated as success — downstream `_add_grid_info_to_json` and `_split_multi_echo_dixon` then ran on a NIfTI that was never actually written. Now the branch logs a warning, deletes the half-written sidecar JSON, and returns False so the caller stops the pipeline for that series. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Multi-echo series (e.g. Philips mDIX quant) put M echoes at each of N slice positions inside a single ImageType sub-group. Sorted by InstanceNumber, consecutive DICOMs then share ImagePositionPatient, so the direction test computed `0 / 0 = NaN` and `np.allclose(NaN, prev)` false-triggered a stack split — one physical acquisition would be shattered into arbitrary chunks. New helper `_split_by_echo_numbers` picks the smallest `EchoNumbers` bucket for stack detection so each position appears exactly once, then re-attaches the remaining echoes to the detected stacks by `ImagePositionPatient`. Also skips the direction update whenever `norm(delta) < 1e-6`, guarding against residual duplicates and genuinely repeated slices. No-op for single-echo series (0 or 1 distinct EchoNumbers value). Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

`_from_dicom_to_nii` only tried to split multi-echo output when the BIDS `part` entity resolved to "magnitude" or "phase". Any 4-D volume whose part was mapped to something else via `dixon_mapping` / `parts_mapping` (or none at all) escaped the split and sat on disk as an opaque 4-D file. Since `_split_multi_echo_dixon` is already a no-op on 3-D input, drop the gate and always try to split. Also switch the frame/TE iteration in `_split_multi_echo_dixon` to `zip_strict`. `tes` is padded to `n_echo` upstream, so `zip` is currently safe — but `zip_strict` makes the invariant explicit and traps future regressions instead of silently truncating. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

The first `yield dcm_dirs` handed back the raw argument, which is a `list[Path]` when the caller passed a list. Downstream code (in `extract_dicom_folder`) then did `str(dicom_path).endswith(".pkl")` on that list and eventually `_read_dicom_files(list)` — the list-repr string match ate the guard and `.rglob` blew up with `AttributeError`. Iterate the normalised `dcm_dirs` list instead, so we always yield `Path` instances one at a time. Behaviour for the common single-Path case is unchanged. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Previously `_read_dicom_files` handed every file under a source folder to `pydicom.dcmread(force=True)` — README.txt, .DS_Store, thumbnail PNGs, JSON sidecars from earlier extractions, and so on. Each dcmread cost real time on top-level scanner exports that mix DICOMs with reports and metadata. New helper `_looks_like_dicom(path)`: 1. Skips files with obviously non-DICOM extensions (json/txt/md/pdf, image formats, csv/tsv/yaml/html/xml, already-extracted NIfTI / nrrd). 2. Reads the first 132 bytes and accepts the standard "DICM" magic at offset 128. 3. Falls back to accepting extensionless files (and .dcm / .ima / .dicom), matching the previous permissive behaviour for Philips/Siemens exports that ship as `IM000001` and skip the preamble. Archive extensions (`.zip` / `.tar` / …) are intentionally NOT in the block list because top-level zips are unpacked by `extract_dicom_folder` before `_read_dicom_files` sees them, and any residual archive would fail the DICM-magic check anyway. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

`_inc_key` incremented its key while `path_exists(keys)` remained True with no upper bound. A mis-configured `path_exists` callback (e.g. one that never sees the mutation because it captures stale state) would spin forever. Add a hard cap and raise `RuntimeError` with the offending key value instead, so the mistake surfaces immediately. 10 000 tries covers `-a` through `-mnw` letter suffixes — well beyond any realistic filename-collision run. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

The previous one-liner claimed "Defaults to 1 (sequential)" but did not mention that `n_cpu=None` — a common thing to try when reaching for "use all cores" — hits the parallel path with `max_workers=None`, meaning `min(32, os.cpu_count() + 4)` Python threads. Spell all three cases out (1 / >1 / None) and note that this workload is I/O-heavy. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Two related plane-detection tweaks: 1. The real caller inside `extract_keys_from_json` was passing `hires_threshold=1` while the docstring / function default advertised 0.8. Push the actual call to 0.8 so the behaviour matches the docs (and so submillimetre isotropic MRI still resolves cleanly after the multi-echo pre-filter took the "everything clamps to 1.0" collapse out of the equation). 2. The bare `except Exception: return None` at the end of `get_plane_dicom` swallowed every failure silently. A follow-up `acq-None` filename would then be unattributable. Log a warning with the exception type + message before returning None, guarding the log import itself so the fallback still returns None cleanly if the logger isn't importable. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Several common series descriptions were misclassified because the greedy ``.*t2.*`` / ``.*t1.*`` catch-alls matched first (first-match wins). Add specific patterns BEFORE them: * Quantitative maps: MP2RAGE, T2*, R2*, T1map, T2map, T2starmap. * SWI (susceptibility weighted). * Multi-echo VIBE variants for both NAKO Siemens (`ME_vibe_fatquant_*`) and the generic mevibe / fatquant labels. * DIXON auxiliary maps: fatfrac, PDFF, GE Ideal. * Philips localizer aliases the existing pilot/scout entries missed: survey, refscan, smartexam. Also add an explicit ``.*mdixon.*`` alongside the existing ``.*mdix.*`` so both spellings of the Philips protocol name resolve to "dixon". No behaviour change for series that were already correctly classified. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

…ypart Three coupled additions so `extract_keys_from_json` stops raising on non-CT/PET/MR/XA modalities. * `bids_constants.formats`: add `xray`, `us`, `nm`, `sc`, `photo`, `endoscopy`, `rtimage`, `rtstruct`, `rtdose`, `rtplan`, `ot`. Without these entries, `BIDS_FILE.get_changed_bids` refuses the new formats under strict mode. * New modality branches in `extract_keys_from_json`. Instead of `raise NotImplementedError`, dispatch: - CR/DX/RG/PX/IO/MG → `xray` - US → `us` - NM → `nm` (Radiopharmaceutical → `ce` when set) - SC → `sc` - OP/XC → `photo` - ES → `endoscopy` - RTIMAGE/RTSTRUCT/RTDOSE/RTPLAN → matching lowercase format - OT → `ot` - unknown → warn and fall back to `modality.lower()` so the pipeline can complete instead of aborting mid-tree. * Two helpers extract the DICOM tags that actually differentiate these files in a session: - `_apply_view_keys` lifts `ViewPosition` into `acq` (only overwriting a None/iso plane) and `ImageLaterality`/`Laterality` into `part`, with ophthalmic OS/OD normalised to L/R. - `_apply_bodypart_key` lifts `BodyPartExamined` into `desc` when empty, matching common values (ABDOMEN, PELVIS, CHEST, HEAD…). Called from CT, PET, and every new modality branch so files that only differ by view (MG four-view screening), laterality (ophthalmic), tracer (NM), or body region no longer collapse onto identical BIDS filenames. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

github-actions · 2026-09-15T08:27:07Z

Benchmark comparison

Speed (wall time per call)

baseline 68adb2a vs head 68adb2a · python 3.11.16 · 5 repeats + 1 warmup

Showing the 5 most-changed measurements per case; the rest are collapsed. Rows with a baseline below 3 ms are tagged (noise) — runner jitter on those swamps any real change.

`ct_3d` — shape (73, 47, 73)

Measurement	baseline ms (median ±½·range)	head ms (median ±½·range)	Δ %	p
`poi_calc_poi_from_subreg_vert`	10.27 ±0.10	10.63 ±0.40	+3.5%	0.037
`poi_calc_centroids_nocrop`	5.43 ±0.04	5.58 ±0.02	+2.7%	0.000
`nii_load_img`	12.72 ±0.09	12.86 ±0.07	+1.1%	0.228
`nii_save`	10.57 ±0.06	10.68 ±0.08	+1.0%	0.087
`nii_dilate_msk`	101.18 ±0.12	102.20 ±0.23	+1.0%	0.000

… 31 more measurements

Measurement	baseline ms (median ±½·range)	head ms (median ±½·range)	Δ %	p
`nii_fill_holes`	4.48 ±0.03	4.46 ±0.08	-0.4%	0.616
`nii_rescale`	42.85 ±0.23	42.68 ±0.98	-0.4%	0.393
`nii_filter_connected_components`	4.16 ±0.03	4.15 ±0.02	-0.2%	0.542
`nii_erode_msk`	6.03 ±0.03	6.04 ±0.05	+0.1%	0.877
`nii_resample_from_to`	44.00 ±0.16	44.04 ±0.22	+0.1%	0.118
`metric_voxels`	250463.00 ±0.00	250463.00 ±0.00	+0.0%	—
`metric_labels`	3.00 ±0.00	3.00 ±0.00	+0.0%	—
`metric_foreground_pct`	14.51 ±0.00	14.51 ±0.00	+0.0%	—
`nii_set_dtype`	0.41 ±0.01	0.37 ±0.02	-9.6% (noise)	0.050
`poi_reorient`	0.30 ±0.02	0.31 ±0.01	+5.9% (noise)	0.841
`poi_rescale`	0.18 ±0.01	0.17 ±0.01	-5.5% (noise)	0.467
`nii_apply_crop`	0.50 ±0.01	0.52 ±0.02	+3.3% (noise)	0.380
`nii_compute_crop`	0.28 ±0.01	0.28 ±0.00	-2.4% (noise)	0.484
`poi_resample_from_to`	0.51 ±0.01	0.52 ±0.01	+2.3% (noise)	0.117
`nii_center_of_masses`	1.54 ±0.02	1.51 ±0.01	-1.9% (noise)	0.054
`poi_to_global`	0.16 ±0.01	0.16 ±0.01	-1.7% (noise)	0.235
`nii_get_connected_components`	2.41 ±0.03	2.45 ±0.01	+1.5% (noise)	0.316
`nii_pad_to`	0.53 ±0.01	0.52 ±0.02	-1.4% (noise)	0.473
`nii_rescale_seg`	2.72 ±0.06	2.75 ±0.14	+1.1% (noise)	0.337
`poi_local_to_global_arr`	0.27 ±0.01	0.27 ±0.01	-1.0% (noise)	0.877
`nii_reorient`	0.64 ±0.03	0.64 ±0.02	-1.0% (noise)	0.496
`poi_load`	0.77 ±0.01	0.78 ±0.04	+0.9% (noise)	0.299
`nii_extract_label`	0.29 ±0.02	0.30 ±0.01	+0.9% (noise)	0.524
`nii_map_labels`	1.01 ±0.01	1.02 ±0.01	+0.8% (noise)	0.330
`poi_save`	0.30 ±0.02	0.30 ±0.02	-0.7% (noise)	0.952
`nii_load_seg`	2.36 ±0.03	2.38 ±0.03	+0.6% (noise)	0.453
`nii_unique`	0.83 ±0.01	0.83 ±0.01	-0.6% (noise)	0.177
`nii_get_array`	0.04 ±0.00	0.04 ±0.00	+0.6% (noise)	0.783
`poi_calc_centroids`	2.06 ±0.03	2.07 ±0.01	+0.4% (noise)	0.667
`poi_map_labels`	0.14 ±0.01	0.14 ±0.01	+0.4% (noise)	0.996
`nii_volumes`	1.64 ±0.01	1.63 ±0.02	-0.2% (noise)	0.523

`ct_2d` — shape (73, 47, 1)

Measurement	baseline ms (median ±½·range)	head ms (median ±½·range)	Δ %	p
`nii_rescale`	5.11 ±0.06	5.15 ±0.05	+0.8%	0.630
`nii_resample_from_to`	6.54 ±0.05	6.52 ±0.07	-0.3%	0.857
`nii_dilate_msk`	11.66 ±0.04	11.67 ±0.24	+0.1%	0.384
`metric_voxels`	3431.00 ±0.00	3431.00 ±0.00	+0.0%	—
`metric_labels`	3.00 ±0.00	3.00 ±0.00	+0.0%	—

… 30 more measurements

Measurement	baseline ms (median ±½·range)	head ms (median ±½·range)	Δ %	p
`metric_foreground_pct`	37.60 ±0.00	37.60 ±0.00	+0.0%	—
`nii_unique`	0.08 ±0.00	0.09 ±0.01	+10.8% (noise)	0.062
`nii_set_dtype`	0.31 ±0.01	0.29 ±0.01	-7.7% (noise)	0.028
`nii_volumes`	0.20 ±0.01	0.21 ±0.00	+7.4% (noise)	0.105
`poi_reorient`	0.34 ±0.01	0.32 ±0.01	-5.1% (noise)	0.262
`nii_center_of_masses`	0.13 ±0.01	0.14 ±0.01	+4.9% (noise)	0.095
`poi_map_labels`	0.16 ±0.01	0.15 ±0.01	-3.7% (noise)	0.493
`nii_apply_crop`	0.50 ±0.02	0.48 ±0.02	-3.6% (noise)	0.240
`nii_compute_crop`	0.10 ±0.01	0.10 ±0.01	+3.5% (noise)	0.301
`nii_extract_label`	0.29 ±0.02	0.30 ±0.01	+3.3% (noise)	0.961
`nii_reorient`	0.62 ±0.02	0.60 ±0.02	-3.2% (noise)	0.175
`poi_local_to_global_arr`	0.30 ±0.01	0.29 ±0.01	-3.0% (noise)	0.339
`nii_load_img`	1.14 ±0.01	1.17 ±0.02	+2.7% (noise)	0.027
`nii_pad_to`	0.41 ±0.02	0.40 ±0.01	-2.4% (noise)	0.145
`poi_calc_centroids`	0.68 ±0.01	0.70 ±0.01	+2.4% (noise)	0.042
`poi_resample_from_to`	0.55 ±0.01	0.56 ±0.01	+1.6% (noise)	0.594
`nii_get_array`	0.01 ±0.00	0.01 ±0.00	+1.5% (noise)	0.384
`nii_rescale_seg`	0.56 ±0.02	0.57 ±0.02	+1.5% (noise)	0.833
`nii_load_seg`	1.18 ±0.03	1.20 ±0.03	+1.5% (noise)	0.465
`poi_rescale`	0.19 ±0.01	0.20 ±0.01	+1.4% (noise)	0.671
`nii_save`	0.99 ±0.02	1.01 ±0.03	+1.4% (noise)	0.762
`poi_calc_centroids_nocrop`	0.86 ±0.02	0.87 ±0.01	+1.2% (noise)	0.476
`nii_fill_holes`	0.64 ±0.01	0.64 ±0.02	+1.1% (noise)	0.829
`poi_to_global`	0.17 ±0.01	0.17 ±0.01	+0.8% (noise)	0.215
`nii_map_labels`	0.35 ±0.01	0.35 ±0.01	-0.7% (noise)	0.907
`nii_get_connected_components`	0.48 ±0.01	0.48 ±0.02	+0.7% (noise)	0.363
`nii_erode_msk`	0.76 ±0.03	0.76 ±0.01	+0.6% (noise)	0.900
`poi_save`	0.34 ±0.01	0.34 ±0.01	-0.3% (noise)	0.889
`poi_load`	0.81 ±0.01	0.81 ±0.02	+0.2% (noise)	0.903
`nii_filter_connected_components`	0.62 ±0.03	0.62 ±0.02	+0.2% (noise)	0.652

`mri_3d` — shape (68, 52, 67)

Measurement	baseline ms (median ±½·range)	head ms (median ±½·range)	Δ %	p
`nii_filter_connected_components`	4.36 ±0.02	4.46 ±0.06	+2.3%	0.012
`poi_calc_centroids_nocrop`	11.49 ±0.04	11.28 ±0.07	-1.8%	0.004
`nii_erode_msk`	7.60 ±0.04	7.67 ±0.04	+0.9%	0.040
`nii_resample_from_to`	41.57 ±0.62	41.85 ±0.98	+0.7%	0.299
`nii_rescale`	40.09 ±0.72	40.35 ±0.97	+0.6%	0.473

… 31 more measurements

Measurement	baseline ms (median ±½·range)	head ms (median ±½·range)	Δ %	p
`poi_calc_poi_from_subreg_vert`	9.52 ±0.06	9.56 ±0.05	+0.5%	0.100
`nii_save`	8.50 ±0.07	8.54 ±0.07	+0.5%	0.267
`nii_fill_holes`	5.49 ±0.03	5.51 ±0.02	+0.4%	0.330
`nii_load_img`	10.15 ±0.04	10.17 ±0.06	+0.2%	0.994
`nii_dilate_msk`	132.10 ±0.32	132.26 ±0.53	+0.1%	0.671
`metric_voxels`	236912.00 ±0.00	236912.00 ±0.00	+0.0%	—
`metric_labels`	8.00 ±0.00	8.00 ±0.00	+0.0%	—
`metric_foreground_pct`	18.21 ±0.00	18.21 ±0.00	+0.0%	—
`nii_get_array`	0.04 ±0.00	0.05 ±0.00	+15.4% (noise)	0.000
`nii_pad_to`	0.55 ±0.01	0.63 ±0.01	+14.2% (noise)	0.000
`nii_extract_label`	0.35 ±0.01	0.38 ±0.02	+9.4% (noise)	0.017
`poi_local_to_global_arr`	0.29 ±0.01	0.32 ±0.01	+9.2% (noise)	0.017
`nii_volumes`	1.84 ±0.02	1.68 ±0.01	-8.5% (noise)	0.000
`nii_center_of_masses`	1.88 ±0.01	1.73 ±0.04	-8.3% (noise)	0.001
`nii_set_dtype`	0.44 ±0.01	0.47 ±0.02	+7.4% (noise)	0.027
`poi_to_global`	0.20 ±0.00	0.21 ±0.01	+4.3% (noise)	0.035
`poi_rescale`	0.20 ±0.00	0.21 ±0.02	+4.3% (noise)	0.121
`poi_resample_from_to`	0.58 ±0.04	0.60 ±0.01	+3.8% (noise)	0.927
`nii_map_labels`	1.04 ±0.01	1.07 ±0.01	+3.7% (noise)	0.003
`nii_get_connected_components`	2.32 ±0.03	2.38 ±0.04	+2.5% (noise)	0.031
`poi_map_labels`	0.20 ±0.01	0.20 ±0.01	-2.5% (noise)	0.224
`nii_unique`	0.85 ±0.02	0.86 ±0.02	+1.9% (noise)	0.038
`nii_compute_crop`	0.30 ±0.02	0.30 ±0.02	-1.4% (noise)	0.793
`poi_calc_centroids`	2.28 ±0.01	2.31 ±0.02	+1.2% (noise)	0.041
`nii_load_seg`	2.32 ±0.02	2.34 ±0.11	+1.0% (noise)	0.240
`poi_load`	1.78 ±0.01	1.79 ±0.00	+0.4% (noise)	0.029
`nii_apply_crop`	0.53 ±0.03	0.54 ±0.00	+0.3% (noise)	0.589
`poi_reorient`	0.36 ±0.01	0.36 ±0.01	+0.2% (noise)	0.256
`nii_reorient`	0.67 ±0.01	0.67 ±0.01	+0.1% (noise)	0.975
`nii_rescale_seg`	2.62 ±0.04	2.62 ±0.07	-0.0% (noise)	0.704
`poi_save`	0.40 ±0.02	0.40 ±0.02	-0.0% (noise)	0.683

`mri_2d` — shape (68, 52, 1)

Measurement	baseline ms (median ±½·range)	head ms (median ±½·range)	Δ %	p
`nii_resample_from_to`	6.53 ±0.06	6.62 ±0.07	+1.3%	0.198
`nii_rescale`	5.14 ±0.05	5.18 ±0.02	+0.7%	0.235
`nii_dilate_msk`	15.46 ±0.09	15.53 ±0.07	+0.4%	0.229
`metric_voxels`	3536.00 ±0.00	3536.00 ±0.00	+0.0%	—
`metric_labels`	7.00 ±0.00	7.00 ±0.00	+0.0%	—

… 30 more measurements

Measurement	baseline ms (median ±½·range)	head ms (median ±½·range)	Δ %	p
`metric_foreground_pct`	50.17 ±0.00	50.17 ±0.00	+0.0%	—
`poi_to_global`	0.21 ±0.00	0.23 ±0.01	+9.2% (noise)	0.026
`nii_extract_label`	0.33 ±0.02	0.30 ±0.01	-9.0% (noise)	0.060
`nii_get_array`	0.01 ±0.00	0.01 ±0.00	+8.5% (noise)	0.079
`poi_rescale`	0.22 ±0.01	0.21 ±0.01	-8.1% (noise)	0.011
`poi_save`	0.38 ±0.00	0.41 ±0.03	+7.6% (noise)	0.093
`nii_get_connected_components`	0.46 ±0.02	0.49 ±0.01	+7.6% (noise)	0.065
`nii_pad_to`	0.41 ±0.02	0.39 ±0.01	-4.5% (noise)	0.099
`nii_unique`	0.08 ±0.00	0.09 ±0.01	+4.5% (noise)	0.181
`nii_center_of_masses`	0.31 ±0.01	0.30 ±0.00	-4.3% (noise)	0.059
`nii_load_seg`	1.17 ±0.02	1.22 ±0.03	+4.2% (noise)	0.004
`nii_rescale_seg`	0.58 ±0.02	0.56 ±0.01	-4.0% (noise)	0.174
`poi_local_to_global_arr`	0.30 ±0.01	0.31 ±0.02	+3.9% (noise)	0.283
`nii_reorient`	0.63 ±0.01	0.61 ±0.03	-3.1% (noise)	0.065
`nii_load_img`	1.11 ±0.01	1.14 ±0.03	+3.1% (noise)	0.030
`nii_volumes`	0.35 ±0.01	0.36 ±0.01	+1.9% (noise)	0.758
`nii_apply_crop`	0.48 ±0.01	0.47 ±0.01	-1.4% (noise)	0.411
`poi_load`	1.58 ±0.02	1.60 ±0.02	+1.4% (noise)	0.308
`nii_compute_crop`	0.10 ±0.00	0.10 ±0.00	-1.3% (noise)	0.411
`nii_set_dtype`	0.30 ±0.02	0.29 ±0.01	-1.3% (noise)	0.137
`poi_map_labels`	0.19 ±0.01	0.19 ±0.01	+1.2% (noise)	0.703
`nii_fill_holes`	0.80 ±0.02	0.79 ±0.02	-1.2% (noise)	0.913
`nii_filter_connected_components`	0.65 ±0.06	0.64 ±0.01	-1.1% (noise)	0.398
`poi_resample_from_to`	0.59 ±0.02	0.60 ±0.02	+0.7% (noise)	0.985
`poi_calc_centroids`	0.87 ±0.01	0.86 ±0.02	-0.7% (noise)	0.547
`nii_save`	0.99 ±0.01	0.98 ±0.02	-0.7% (noise)	0.872
`nii_map_labels`	0.37 ±0.01	0.36 ±0.01	-0.5% (noise)	0.606
`nii_erode_msk`	0.97 ±0.00	0.97 ±0.01	-0.4% (noise)	0.183
`poi_reorient`	0.35 ±0.01	0.35 ±0.01	-0.3% (noise)	0.358
`poi_calc_centroids_nocrop`	1.09 ±0.01	1.09 ±0.02	+0.1% (noise)	0.677

Gate: a measurement fails when the baseline is ≥ 1 ms, the median grows by ≥ 50%, and Welch's t-test gives p < 0.05. metric_* rows are context only.

Memory (peak RSS growth per call)

baseline 68adb2a vs head 68adb2a · python 3.11.16 · 5 repeats + 1 warmup · sampler proc-status, isolation fork · measurement floor ≈ 0.95 MiB

Showing the 5 most-changed measurements per case; the rest are collapsed. Rows with a baseline below 3 MiB are tagged (noise) — runner jitter on those swamps any real change.

`ct_3d` — shape (73, 47, 73)

Measurement	baseline MiB (median ±½·range)	head MiB (median ±½·range)	Δ %	p
`poi_resample_from_to`	5.59 ±0.00	5.49 ±0.00	-1.7%	—
`nii_reorient`	6.03 ±0.00	5.93 ±0.00	-1.6%	—
`nii_load_seg`	6.13 ±0.00	6.04 ±0.00	-1.5%	—
`nii_resample_from_to`	6.61 ±0.00	6.51 ±0.00	-1.5%	—
`nii_rescale`	6.48 ±0.00	6.39 ±0.00	-1.4%	—

… 31 more measurements

Measurement	baseline MiB (median ±½·range)	head MiB (median ±½·range)	Δ %	p
`nii_load_img`	3.62 ±0.00	3.58 ±0.00	-1.2%	—
`nii_dilate_msk`	6.12 ±0.00	6.05 ±0.00	-1.1%	—
`nii_set_dtype`	5.43 ±0.00	5.37 ±0.00	-1.1%	—
`nii_rescale_seg`	6.25 ±0.00	6.19 ±0.00	-1.1%	—
`nii_pad_to`	5.55 ±0.00	5.49 ±0.00	-1.1%	—
`nii_extract_label`	5.55 ±0.00	5.50 ±0.00	-1.1%	—
`nii_map_labels`	5.62 ±0.00	5.56 ±0.00	-1.0%	—
`nii_apply_crop`	5.99 ±0.00	5.93 ±0.00	-1.0%	—
`poi_calc_centroids_nocrop`	6.11 ±0.00	6.05 ±0.00	-1.0%	—
`nii_erode_msk`	6.12 ±0.00	6.06 ±0.00	-1.0%	—
`poi_calc_centroids`	6.24 ±0.00	6.18 ±0.00	-0.9%	—
`nii_get_connected_components`	6.74 ±0.00	6.68 ±0.00	-0.9%	—
`nii_fill_holes`	6.47 ±0.00	6.41 ±0.00	-0.8%	—
`nii_filter_connected_components`	7.05 ±0.00	7.00 ±0.00	-0.8%	—
`poi_calc_poi_from_subreg_vert`	7.36 ±0.00	7.30 ±0.00	-0.8%	—
`nii_compute_crop`	3.02 ±0.00	3.02 ±0.00	-0.3%	—
`poi_local_to_global_arr`	3.08 ±0.00	3.07 ±0.00	-0.3%	—
`poi_to_global`	3.21 ±0.00	3.20 ±0.00	-0.2%	—
`nii_center_of_masses`	3.27 ±0.00	3.27 ±0.00	-0.2%	—
`nii_volumes`	3.59 ±0.00	3.58 ±0.00	-0.2%	—
`nii_save`	3.90 ±0.00	3.89 ±0.00	-0.2%	—
`poi_reorient`	3.96 ±0.00	3.95 ±0.00	-0.2%	—
`metric_voxels`	250463.00 ±0.00	250463.00 ±0.00	+0.0%	—
`metric_labels`	3.00 ±0.00	3.00 ±0.00	+0.0%	—
`metric_foreground_pct`	14.51 ±0.00	14.51 ±0.00	+0.0%	—
`poi_save`	1.53 ±0.00	1.52 ±0.00	-0.5% (noise)	—
`poi_load`	1.68 ±0.00	1.67 ±0.00	-0.5% (noise)	—
`nii_get_array`	1.71 ±0.00	1.70 ±0.00	-0.5% (noise)	—
`poi_map_labels`	1.84 ±0.00	1.83 ±0.00	-0.4% (noise)	—
`poi_rescale`	2.67 ±0.00	2.66 ±0.00	-0.3% (noise)	—
`nii_unique`	2.96 ±0.00	2.95 ±0.00	-0.3% (noise)	—

`ct_2d` — shape (73, 47, 1)

Measurement	baseline MiB (median ±½·range)	head MiB (median ±½·range)	Δ %	p
`poi_resample_from_to`	5.53 ±0.00	5.43 ±0.00	-1.8%	—
`nii_reorient`	5.97 ±0.00	5.87 ±0.00	-1.6%	—
`nii_load_seg`	6.07 ±0.00	5.98 ±0.00	-1.5%	—
`nii_resample_from_to`	6.55 ±0.00	6.45 ±0.00	-1.5%	—
`nii_rescale`	6.42 ±0.00	6.32 ±0.00	-1.5%	—

… 30 more measurements

Measurement	baseline MiB (median ±½·range)	head MiB (median ±½·range)	Δ %	p
`nii_load_img`	3.56 ±0.00	3.52 ±0.00	-1.2%	—
`nii_dilate_msk`	6.05 ±0.00	5.98 ±0.00	-1.2%	—
`nii_set_dtype`	5.37 ±0.00	5.31 ±0.00	-1.1%	—
`nii_rescale_seg`	6.19 ±0.00	6.12 ±0.00	-1.1%	—
`nii_pad_to`	5.49 ±0.00	5.43 ±0.00	-1.1%	—
`nii_extract_label`	5.49 ±0.00	5.43 ±0.00	-1.1%	—
`nii_map_labels`	5.55 ±0.00	5.50 ±0.00	-1.1%	—
`nii_apply_crop`	5.93 ±0.00	5.87 ±0.00	-1.0%	—
`nii_erode_msk`	6.06 ±0.00	6.00 ±0.00	-1.0%	—
`poi_calc_centroids_nocrop`	6.11 ±0.00	6.05 ±0.00	-1.0%	—
`poi_calc_centroids`	6.18 ±0.00	6.12 ±0.00	-0.9%	—
`nii_get_connected_components`	6.68 ±0.00	6.62 ±0.00	-0.9%	—
`nii_fill_holes`	6.41 ±0.00	6.35 ±0.00	-0.9%	—
`nii_filter_connected_components`	6.99 ±0.00	6.93 ±0.00	-0.8%	—
`poi_local_to_global_arr`	3.02 ±0.00	3.01 ±0.00	-0.3%	—
`poi_to_global`	3.14 ±0.00	3.14 ±0.00	-0.2%	—
`nii_center_of_masses`	3.21 ±0.00	3.20 ±0.00	-0.2%	—
`nii_volumes`	3.52 ±0.00	3.52 ±0.00	-0.2%	—
`nii_save`	3.84 ±0.00	3.83 ±0.00	-0.2%	—
`poi_reorient`	3.89 ±0.00	3.89 ±0.00	-0.2%	—
`metric_voxels`	3431.00 ±0.00	3431.00 ±0.00	+0.0%	—
`metric_labels`	3.00 ±0.00	3.00 ±0.00	+0.0%	—
`metric_foreground_pct`	37.60 ±0.00	37.60 ±0.00	+0.0%	—
`poi_save`	1.46 ±0.00	1.46 ±0.00	-0.5% (noise)	—
`poi_load`	1.61 ±0.00	1.61 ±0.00	-0.5% (noise)	—
`nii_get_array`	1.65 ±0.00	1.64 ±0.00	-0.5% (noise)	—
`poi_map_labels`	1.77 ±0.00	1.77 ±0.00	-0.4% (noise)	—
`poi_rescale`	2.61 ±0.00	2.60 ±0.00	-0.3% (noise)	—
`nii_unique`	2.90 ±0.00	2.89 ±0.00	-0.3% (noise)	—
`nii_compute_crop`	2.96 ±0.00	2.95 ±0.00	-0.3% (noise)	—

`mri_3d` — shape (68, 52, 67)

Measurement	baseline MiB (median ±½·range)	head MiB (median ±½·range)	Δ %	p
`poi_resample_from_to`	5.53 ±0.00	5.43 ±0.00	-1.8%	—
`nii_reorient`	5.97 ±0.00	5.87 ±0.00	-1.6%	—
`nii_load_seg`	6.07 ±0.00	5.98 ±0.00	-1.5%	—
`nii_resample_from_to`	6.55 ±0.00	6.45 ±0.00	-1.5%	—
`nii_rescale`	6.42 ±0.00	6.32 ±0.00	-1.5%	—

… 31 more measurements

Measurement	baseline MiB (median ±½·range)	head MiB (median ±½·range)	Δ %	p
`nii_load_img`	3.62 ±0.00	3.58 ±0.00	-1.2%	—
`nii_dilate_msk`	6.05 ±0.00	5.98 ±0.00	-1.2%	—
`nii_set_dtype`	5.43 ±0.00	5.37 ±0.00	-1.1%	—
`nii_rescale_seg`	6.19 ±0.00	6.12 ±0.00	-1.1%	—
`nii_pad_to`	5.49 ±0.00	5.43 ±0.00	-1.1%	—
`nii_extract_label`	5.49 ±0.00	5.43 ±0.00	-1.1%	—
`nii_map_labels`	5.55 ±0.00	5.50 ±0.00	-1.1%	—
`nii_apply_crop`	5.93 ±0.00	5.87 ±0.00	-1.0%	—
`poi_calc_centroids_nocrop`	6.05 ±0.00	5.99 ±0.00	-1.0%	—
`nii_erode_msk`	6.06 ±0.00	6.00 ±0.00	-1.0%	—
`poi_calc_centroids`	6.18 ±0.00	6.12 ±0.00	-0.9%	—
`nii_get_connected_components`	6.68 ±0.00	6.62 ±0.00	-0.9%	—
`nii_fill_holes`	6.41 ±0.00	6.35 ±0.00	-0.9%	—
`nii_filter_connected_components`	6.99 ±0.00	6.93 ±0.00	-0.8%	—
`poi_calc_poi_from_subreg_vert`	7.36 ±0.00	7.30 ±0.00	-0.8%	—
`poi_local_to_global_arr`	3.02 ±0.00	3.01 ±0.00	-0.3%	—
`poi_to_global`	3.14 ±0.00	3.14 ±0.00	-0.2%	—
`nii_center_of_masses`	3.21 ±0.00	3.20 ±0.00	-0.2%	—
`nii_volumes`	3.52 ±0.00	3.52 ±0.00	-0.2%	—
`nii_save`	3.84 ±0.00	3.83 ±0.00	-0.2%	—
`poi_reorient`	3.89 ±0.00	3.89 ±0.00	-0.2%	—
`metric_voxels`	236912.00 ±0.00	236912.00 ±0.00	+0.0%	—
`metric_labels`	8.00 ±0.00	8.00 ±0.00	+0.0%	—
`metric_foreground_pct`	18.21 ±0.00	18.21 ±0.00	+0.0%	—
`poi_save`	1.53 ±0.00	1.52 ±0.00	-0.5% (noise)	—
`poi_load`	1.61 ±0.00	1.61 ±0.00	-0.5% (noise)	—
`nii_get_array`	1.65 ±0.00	1.64 ±0.00	-0.5% (noise)	—
`poi_map_labels`	1.77 ±0.00	1.77 ±0.00	-0.4% (noise)	—
`poi_rescale`	2.61 ±0.00	2.60 ±0.00	-0.3% (noise)	—
`nii_unique`	2.90 ±0.00	2.89 ±0.00	-0.3% (noise)	—
`nii_compute_crop`	2.96 ±0.00	2.95 ±0.00	-0.3% (noise)	—

`mri_2d` — shape (68, 52, 1)

Measurement	baseline MiB (median ±½·range)	head MiB (median ±½·range)	Δ %	p
`poi_resample_from_to`	5.53 ±0.00	5.43 ±0.00	-1.8%	—
`nii_reorient`	5.97 ±0.00	5.87 ±0.00	-1.6%	—
`nii_load_seg`	6.07 ±0.00	5.98 ±0.00	-1.5%	—
`nii_resample_from_to`	6.55 ±0.00	6.45 ±0.00	-1.5%	—
`nii_rescale`	6.42 ±0.00	6.32 ±0.00	-1.5%	—

… 30 more measurements

Measurement	baseline MiB (median ±½·range)	head MiB (median ±½·range)	Δ %	p
`nii_load_img`	3.62 ±0.00	3.58 ±0.00	-1.2%	—
`nii_dilate_msk`	6.05 ±0.00	5.98 ±0.00	-1.2%	—
`nii_set_dtype`	5.43 ±0.00	5.37 ±0.00	-1.1%	—
`nii_rescale_seg`	6.19 ±0.00	6.12 ±0.00	-1.1%	—
`nii_pad_to`	5.49 ±0.00	5.43 ±0.00	-1.1%	—
`nii_extract_label`	5.49 ±0.00	5.43 ±0.00	-1.1%	—
`nii_map_labels`	5.55 ±0.00	5.50 ±0.00	-1.1%	—
`nii_apply_crop`	5.93 ±0.00	5.87 ±0.00	-1.0%	—
`nii_erode_msk`	6.06 ±0.00	6.00 ±0.00	-1.0%	—
`poi_calc_centroids_nocrop`	6.11 ±0.00	6.05 ±0.00	-1.0%	—
`poi_calc_centroids`	6.18 ±0.00	6.12 ±0.00	-0.9%	—
`nii_get_connected_components`	6.68 ±0.00	6.62 ±0.00	-0.9%	—
`nii_fill_holes`	6.41 ±0.00	6.35 ±0.00	-0.9%	—
`nii_filter_connected_components`	6.99 ±0.00	6.93 ±0.00	-0.8%	—
`poi_local_to_global_arr`	3.02 ±0.00	3.01 ±0.00	-0.3%	—
`poi_to_global`	3.14 ±0.00	3.14 ±0.00	-0.2%	—
`nii_center_of_masses`	3.21 ±0.00	3.20 ±0.00	-0.2%	—
`nii_volumes`	3.52 ±0.00	3.52 ±0.00	-0.2%	—
`nii_save`	3.84 ±0.00	3.83 ±0.00	-0.2%	—
`poi_reorient`	3.89 ±0.00	3.89 ±0.00	-0.2%	—
`metric_voxels`	3536.00 ±0.00	3536.00 ±0.00	+0.0%	—
`metric_labels`	7.00 ±0.00	7.00 ±0.00	+0.0%	—
`metric_foreground_pct`	50.17 ±0.00	50.17 ±0.00	+0.0%	—
`poi_save`	1.53 ±0.00	1.52 ±0.00	-0.5% (noise)	—
`poi_load`	1.61 ±0.00	1.61 ±0.00	-0.5% (noise)	—
`nii_get_array`	1.65 ±0.00	1.64 ±0.00	-0.5% (noise)	—
`poi_map_labels`	1.77 ±0.00	1.77 ±0.00	-0.4% (noise)	—
`poi_rescale`	2.61 ±0.00	2.60 ±0.00	-0.3% (noise)	—
`nii_unique`	2.90 ±0.00	2.89 ±0.00	-0.3% (noise)	—
`nii_compute_crop`	2.96 ±0.00	2.95 ±0.00	-0.3% (noise)	—

Gate: a measurement fails when the baseline is ≥ 1 MiB, the median grows by ≥ 50%, and Welch's t-test gives p < 0.05. metric_* rows are context only.

robert-graf and others added 27 commits August 26, 2026 14:42

add claude reg info

7f04f69

Merge branch 'Point-Reg' of github.com:Hendrik-code/TPTBox into Point…

d7ef944

…-Reg

Merge branch 'main' of github.com:Hendrik-code/TPTBox

3752050

add smauglab support for internal trainier

d7268a6

edgecases

ea7de02

speed up, split vert and ivd

d0ebb89

update Readme

b12cf44

update nako loader

69fdf16

nako cannonical

4bb6714

add prelimenary scoliose and pelic paramerters

bd99bb6

add NAKO Head scanns

ba8269e

Merge branch 'Point-Reg' of github.com:Hendrik-code/TPTBox into Point…

173af27

…-Reg

Merge branch 'nako-follow-up-export' into Point-Reg

c75ecce

fix multi echo bug

07aac40

Merge branch 'main' of github.com:Hendrik-code/TPTBox

371757e

Merge branch 'main' into Point-Reg

ab7dded

robert-graf closed this Sep 15, 2026

robert-graf deleted the Point-Reg branch September 15, 2026 08:24

robert-graf mentioned this pull request Sep 15, 2026

DICOM extract: multi-echo plane detection, modality fallback, and 9 smaller fixes #153

Merged

1 of 2 tasks

Navigation Menu

Repository navigation

DICOM extract: multi-echo plane detection, modality fallback, and 9 smaller fixes - #152

DICOM extract: multi-echo plane detection, modality fallback, and 9 smaller fixes#152
robert-graf wants to merge 27 commits into
mainfrom
Point-Reg

robert-graf commented Sep 15, 2026

Uh oh!

github-actions Bot commented Sep 15, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

robert-graf commented Sep 15, 2026

Summary

Correctness fixes

Perf / usability

Features

Test plan

Uh oh!

github-actions Bot commented Sep 15, 2026

Benchmark comparison

Speed (wall time per call)

ct_3d — shape (73, 47, 73)

ct_2d — shape (73, 47, 1)

mri_3d — shape (68, 52, 67)

mri_2d — shape (68, 52, 1)

Memory (peak RSS growth per call)

ct_3d — shape (73, 47, 73)

ct_2d — shape (73, 47, 1)

mri_3d — shape (68, 52, 67)

mri_2d — shape (68, 52, 1)

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

`ct_3d` — shape (73, 47, 73)

`ct_2d` — shape (73, 47, 1)

`mri_3d` — shape (68, 52, 67)

`mri_2d` — shape (68, 52, 1)

`ct_3d` — shape (73, 47, 73)

`ct_2d` — shape (73, 47, 1)

`mri_3d` — shape (68, 52, 67)

`mri_2d` — shape (68, 52, 1)