Scientific imaging is a system, not a sequence of tools.

01 · Define the Measurement

Start with what the experiment must establish.

Imaging projects often begin with an available instrument, model, or analysis method. Fovea Lab begins with the claim the experiment must support. What needs to be detected or measured? At what spatial, temporal, or molecular scale? Which errors would change the conclusion, and what level of uncertainty is acceptable?

Those questions determine more than the final metric. They shape acquisition conditions, calibration, metadata, controls, reference data, evaluation criteria, and the operating boundaries of the analysis. Defining the measurement first prevents the system from optimizing an attractive image or benchmark score that does not answer the scientific question.

02 · Design the System

Treat the instrument and computation as one measurement chain.

Information is created, transformed, and sometimes lost across the entire imaging system. Optical configuration affects acquisition. Acquisition affects restoration and segmentation. Data organization affects what can be reproduced. Models depend on the population and conditions in which they are used. Scientific review determines whether the final output is meaningful.

Fovea Lab designs across these boundaries rather than treating optics, hardware, software, data, and analysis as separate projects. Components remain modular, but their interfaces, assumptions, and responsibilities are made explicit. Scientists can inspect and guide the workflow, while software and agents assist within defined technical and scientific boundaries.

03 · Verify in Context

Test each transformation where it can fail.

A pipeline can run successfully while producing the wrong result. Image quality can change across depth or time, metadata can become detached from the data, a model can encounter an unfamiliar population, and a parameter change can alter a downstream measurement without producing an obvious error.

Fovea Lab places checks at the stages where these failures arise and evaluates them against the conditions of the experiment. Inputs, parameters, models, intermediate outputs, approvals, and final measurements remain connected through the run. Results can therefore be reproduced and compared, while unexpected behavior can be traced to the point where it entered the system.

From image to insight

The pipeline behind the insight

  1. QC: manifest validation report

  2. QC: focus / illumination QC map

  3. QC: mask confidence report

  4. QC: % replicating + control summary

Model development & monitoring

The ML stages above are built and kept honest by a separate loop. Each model trains on the output of the step before the node it powers, then deploys back into it.

  1. Curate Evidence Sets
  2. Train & Stress-Test
  3. Validate & Calibrate
  4. Register & Deploy
  5. Monitor & Retrain
  • Correction / restoration model → powers correct, trains on ingest output
  • Segmentation model → powers segment, trains on correct output
  • Representation / embedding model → powers profile, trains on segment output
  • MoA / phenotype model → powers interpret, trains on validate output

What holds it up

Engineered to hold up

A trusted result is more than a model output. It is a pipeline with controls, provenance, reproducibility, and a clear audit trail.

No black box

Built from tools your scientists already trust — CellProfiler, Cellpose, StarDist, BaSiCPy, Pycytominer — plus your own code. Every step is open to inspect or swap.

Reproduce it exactly

Each run pins its code, parameters, container, and model weights. Reproduce a result months later, or re-run the same configuration on new data.

QC at every step

Every stage emits a checkable artifact — image, mask, profile, and metadata QC. Bad runs surface early instead of contaminating the result.

Runs where data lives

One pipeline runs on a workstation, your HPC, or the cloud — no rewrite to move. Your data never has to leave its environment.

Pilot to full screen

The pilot pipeline is the production pipeline — tiled, parallel, resumable. Scale throughput without re-engineering the analysis.

Built for terabytes

Whole-slide and high-content runs hit terabyte scale. The pipeline streams and tiles the data so jobs finish and stay tracked.