From Research Script to Production Pipeline

The problem — Most microscopy analysis starts life as a research script: a notebook run top-to-bottom once, on one machine, by the person who wrote it, on the data it was tuned for. Or it starts as a community model — a Cellpose checkpoint, a ZeroCostDL4Mic notebook, a bioimage.io or napari-hub tool — published to demonstrate a method, not to be depended on. Both are valuable and neither is a pipeline. The gap is everything a research artifact is allowed to skip: pinned dependencies, captured parameters, input QC, error handling, provenance, and validation on your data rather than the author's. Adopting one as-is means inheriting its assumptions silently — the cell sizes it was trained on, the modality it never saw, the batch it was tuned against.

What it is / how it works — Promotion is a sequence of concrete moves, each closing one of those gaps:

• Pin and capture. Freeze code, environment, and model weights to exact versions, and lift every threshold, seed, and tile size out of the notebook into recorded configuration — because configuration debt and data dependencies are where a reproducible run quietly dies. This is the substance of a pipeline you can re-run.
• Wrap with QC and contracts. Add input gates (focus, SNR, format, channel order) and explicit hard requirements, so an out-of-distribution input is rejected instead of silently producing a confident, wrong answer.
• Re-validate on your population. A community model's reported accuracy is a claim about its benchmark. Re-validate against the biological question with a metric chosen for it (Metrics Reloaded) — validation without a ground truth, not the paper's number.
• Emit provenance. Every output carries a manifest: which model, which version, which parameters, on which input — so the result is auditable by someone who wasn't there.

The through-line is that the model code is the part that already works; the engineering is the system around it — exactly why a model isn't the system.

Where it breaks — The general failure is adopting a research artifact's positive claim without its negative constraints — and it bites differently per platform. In Cell Painting, a segmentation model tuned on one cell line under-segments a denser one, and the error propagates invisibly into the feature matrix and the percent-replicating readout. In digital pathology, a notebook validated on one cohort silently degrades on a new scanner or stain lot — a generalization gap the original script never had to face. In light-sheet, a destriping or deconvolution script tuned on a shallow stack distorts structure in deep, low-SNR regions it was never tested on. In live-cell, a tracker that worked on a curated clip loses lineage as imaging interval and density change. The discipline is to treat the research artifact as a candidate component — pinned, gated, and re-validated — never as a finished pipeline, so adopting open work compounds trust instead of importing hidden debt.

References

• Sculley et al. (2015). Hidden Technical Debt in Machine Learning Systems. NeurIPS.
• Varoquaux & Cheplygina (2022). Machine Learning for Medical Imaging: Methodological Failures and Recommendations. npj Digital Medicine.
• Maier-Hein, Reinke et al. (2024). Metrics Reloaded: Recommendations for Image Analysis Validation. Nature Methods.