Label-Free / QPI: a pipeline architecture for trustworthy imaging.
01 · Phenotypic Discovery
Find meaningful biological responses across large experimental spaces.
High-content experiments generate more images than scientists can examine directly, but scale alone does not produce insight. The analysis must preserve the relationship between perturbations, controls, cells, wells, plates, and batches while turning millions of individual observations into comparable phenotypic profiles.
Fovea Lab supports the full path from image correction and cellular segmentation to feature extraction, representation learning, batch correction, and biological readout. Each result remains connected to the conditions and quality checks that produced it, allowing researchers to distinguish reproducible phenotypes from technical variation and investigate rare responses that would otherwise disappear inside population averages.
02 · Space & Time
Follow biological systems as they form, move, and change.
Volumetric and time-resolved imaging reveals processes that cannot be understood from a single field or endpoint. Cells migrate, tissues reorganize, structures develop, and rare events emerge across large image volumes. Extracting those events requires more than processing individual frames: the analysis must remain consistent across depth, time, scale, and changing image quality.
Fovea Lab brings acquisition context, restoration, segmentation, registration, tracking, and quantitative measurement into one inspectable workflow. Researchers can follow an observation from a rendered volume or trajectory back to the source data and processing decisions that created it, while quality checks expose where depth, motion, density, or signal loss may have changed the reliability of the result.
03 · Tissue & Spatial Biology
Connect cellular organization with molecular and pathological context.
Tissue images contain information at several scales at once: cellular morphology, local neighborhoods, larger anatomical structures, molecular measurements, and clinically meaningful regions. These signals are often processed in separate systems, making it difficult to understand how a final classification, spatial association, or tissue-level measurement was formed.
Fovea Lab structures image processing, registration, object detection, spatial measurement, and model interpretation as one traceable analysis. Imaging and molecular information can remain aligned through each transformation, while provenance, cohort-aware validation, and reviewable outputs help researchers examine whether a finding is consistent across samples, scanners, staining conditions, and experimental sites.
Pipeline architectures, module by module.
WHAT IT IS
Phase as measurement
Unstained cells bend the light passing through them, and how much they bend it is a physical quantity rather than a stain intensity.
WHY IT FAILS
Optical variance
Dropping the stain drops its batch drift and leaves the optics: background tilt, defocus, aberration and the medium itself all enter the measurement as signal.
QUALITY GATE
Mass needs its conditions
Whether a dry mass is absolute or relative depends on the medium the cells sat in, and that cannot be recovered afterwards.
GET HELP WITH
Label-free pipelines
Bring a dataset whose numbers shift when the reconstruction settings change; the medium, the unwrapping and the calibration are the first things to measure.
Pipeline backbone — reshapes per modality. Select a module for its sub-steps, tools, and deliverable.
Label-Free / QPI
High confidenceIngest
Raw camera frames, vendor containers and already-reconstructed phase maps are read off the acquisition machine, indexed into an addressable series and written into a chunked, lazily readable store. In QPI this stage does not yet produce images: for off-axis and QLSI acquisitions the file on disk is a fringe pattern, and the phase map an analyst wants does not exist until preprocess runs. Ingest therefore preserves the raw measurement unmodified beside the derived store so it can be re-reconstructed later with better parameters. Series identity, frame ordering, bit depth and byte-level integrity are established here, before any derived artifact is written.
Operations
- Acquisition inventory and series indexingEnumerate the acquisition folder into an addressable index over positions, timepoints, z planes and illumination angles, and reconcile it against the acquisition plan.
- Format detection and reader selectionClassify each input into one of qpformat's three families — raw off-axis holograms, raw QLSI interferograms, ready-phase — and select the matching reader.
- Vendor container parsingParse Tomocube .TCF, an HDF5 container readable with h5py, into OME-TIFF; both the per-image and the overall config sidecars must travel with the folder.
- Chunked store conversionWrite OME-Zarr or qpimage HDF5 with a stable per-frame identity and chunk shapes chosen for the downstream access pattern, not the acquisition order.
- Lazy series bindingBind large hologram folders as lazily evaluated series; an ingest that materializes them fails at plate scale rather than at the first bad frame.
- Raw preservation and checksummingHash every raw file with SHA-256 and archive it unmodified; the derived store references the hash, so a re-reconstruction is provably from the same measurement.
Inputs
- Raw off-axis hologram series (specimen on a fringe carrier)TIFF · zipped TIFF series · HDF5
- QLSI interferograms and already-reconstructed phase or complex fieldsTIFF · Phasics TIFF · qpimage HDF5 · .npy
- Holotomography folders with vendor sidecarsTCF (HDF5) · config.dat · JobParameter.tcp
Outputs
- Chunked, lazily readable QPI store with per-frame identityOME-Zarr · qpimage HDF5 · OME-TIFF
- Ingest manifest and preserved raw archiveParquet · YAML · original vendor formats
QC checks
- Frame-count completeness — Verify frame count against the acquisition plan across angles, z planes, timepoints and positions.
- Saturation screen on the raw frame — Fraction of pixels at the sensor ceiling, computed before any reconstruction; the failure most likely to survive undetected to readout.
- Bit-depth preservation — Confirm the frame carries the detector's native bit depth and was not rescaled to 8-bit by a vendor export path.
- Series ordering verification — Cross-check frame order against timestamps or the vendor index rather than filename sort order.
- Integrity and lazy-read verification — SHA-256 every raw file before writing derived data, and measure peak resident memory to confirm the series is not materialized.
Failure modes it guards against
- Vendor export silently rescaled to 8-bit — Reject series whose bit depth does not match the instrument record; ingest from the acquisition machine, not an export folder.
- Zipped TIFF series whose lexical order is not acquisition order — Derive order from timestamps or the vendor index and flag disagreement with the filename sort.
- Holograms discarded after a vendor phase export — Preserve the raw measurement as a first-class deliverable and make raw retention a written condition of the engagement.
- Vendor sidecar configuration left behind when a folder is copied — Check for both configs before conversion starts and refuse to proceed without them.
- Series materialized into memory at ingest — Bind folders through a lazily evaluated series class and assert peak resident memory in the ingest test.
Tools (8)
QPI-native reader over off-axis hologram, QLSI and ready-phase formats, with lazily evaluated series for folders of many holograms.
HDF5-backed container qpformat loads into; stores phase and amplitude separately with QPI metadata attached.
- tcf_to_ometiffOSS · MIT
Parses Tomocube .TCF into OME-TIFF with OME-XML-compliant metadata across the HT, fluorescence and brightfield channels.
- h5pyLibrary · BSD-3-Clause
HDF5 access layer beneath .TCF, qpimage stores and the HyperSpy hologram variants.
- tifffileLibrary · BSD-3-Clause
Reads raw interferogram TIFF, zipped series and vendor OME-TIFF without a JVM dependency.
- Bio-FormatsOSS · GPL-2.0
Reader for roughly 150 vendor formats where the rig writes a general microscopy format rather than a QPI-native one.
- bioformats2rawOSS · GPL-2.0
Batch converter from vendor formats to OME-Zarr, preserving OME metadata.
- ome-zarr-pyOSS · BSD-2-Clause
Reference NGFF reader and writer; produces the multiscales and axes attributes.
Metadata captured
| Field | Type | Example | Standard |
|---|---|---|---|
| acquisition_kind | string | off_axis_hologram | — |
| source_format_family | string | fmts_raw_oah | — |
| frame_index | object | {'position': 12, 'timepoint': 4, 'angle': 37} | OME-NGFF |
| frames_expected_vs_found | object | {'expected': 4800, 'found': 4712} | — |
| series_order_source | string | acquisition_timestamp | — |
| bit_depth | int (bit) | 12 | OME-XML |
| saturated_pixel_fraction | float (fraction) | 0.0004 | — |
| raw_checksum_sha256 | string | 9c1e...4b | — |
| vendor_sidecar_present | object | {'per_image_config': True, 'overall_config': False} | — |
You get
- Chunked QPI store · OME-Zarr / qpimage HDF5
- Ingest manifest and raw archive · Parquet/YAML + original formats
References
- Towards End-to-End Automated Microscopy Control using Holotomography (Zwirnmann, Knobbe & Haddadin, IFAC-PapersOnLine 56(2):6477-6483) paper
- qpformat documentation — the format-proliferation problem and the lazy series design docs
- Compression benchmarking of holotomography data using the OME-Zarr storage format (Lee et al., arXiv 2503.18037) preprint