Before the stain
Decide which wells are worth a stain, a passage or a sequencing run, from the photo you already take.
RenderTheia by Prodius · Organoid quality control
Upload one bright-field photo. Theia finds each organoid, measures it, and answers keep, discard or cannot decide, with the reason.
Render, not a micrographBuilt for
Keep, discard or cannot decide, per organoid and per image, with the value that decided it. A blurry or empty image gets cannot decide, never a guess.
01
Which wells go on to a stain or a sequencing run is decided at the eyepiece, and the decision is not recorded. Organoid cultures vary from batch to batch, so the same well can get a different call on a different day. [8]
02
Published organoid tools report area, circularity, solidity and counts. None of them returns keep or discard for an organoid, and none refuses an image it cannot measure. [1, 2]
03
Focus measures depend on the microscope and the content of the image, so a blur threshold has to be calibrated per instrument. No published cut exists for bright-field organoid images. [6, 5]
04
FDA's 2025 roadmap names organoids among the alternatives to animal testing. A culture that fails quality control after the expensive step costs the study, not just the well. [10]
One image, step by step
1Upload
stored once under its hash
2Gate
contrast, then focus
3Find
each organoid, split where two touch
4Measure
size, shape, edge, overlap
5Call
keep, discard, cannot decide
Render, not a micrographPNG, JPEG or TIFF, 8 or 16 bit. The file is stored once under its hash and never changed. Give the pixel size if you know it; without it the size rules abstain.
Write-once, content-addressed
An empty well or a blurry image stops here with cannot decide and the measurement that stopped it. Noise never gets a keep or discard.
Contrast gate, focus gate
Objects are found, split where two touch, and measured: equivalent diameter, circularity, solidity, whether the frame cuts them, whether boxes overlap.
Size, shape, edge, overlap
Keep, discard or cannot decide, per organoid and for the image, with the rule and the value that decided it and the pipeline version that produced it.
A person can overrule any call
Review and audit
A person overrules any call with a reason; the pipeline's call stays visible and the log is append-only.
Three answers
Keep, discard or cannot decide, per organoid and per image, with the rule and the measured value.
Recorded pipeline
A versioned spec with a hash: gates, segmentation, measurements and rules, each threshold with its source.
Image store
Every upload stored once under its sha256, read-only, with its pixel size.
Regulatory pull
The FDA Modernization Act 2.0 removed the requirement in 2022. FDA's 2025 roadmap names organoids, organ chips and AI models as alternatives. More organoid studies means more wells to judge. [9, 10]
Organoids everywhere
Since the first adult stem cell organoids, bright-field imaging has been the daily check on a culture. The input to a QC call exists in every lab; the call does not. [7, 8]
Open data
Thousands of annotated organoid images have been released with permissive licenses, enough to train and test a detector on public data before any lab shares a single image. [3, 2, 1]
Honest rules
Pipeline v0.1 is classical: a threshold, a focus measure, shape rules. Every threshold carries its source or the word ours, and the version hash pins every number on screen. [4, 5]
Five stages, from a written-down pipeline to rules per organoid type.
01 · We are here
Gates, segmentation and shape rules with every threshold sourced. Runs on example data and public images.
02 · Next
A small detector trained only on permissively licensed public images, shipped as a new recorded version with a held-out test.
03 · Later
Real images from one lab, with their own pixel size, instrument and organoid type.
04 · Later
Size and shape floors set per type with the lab, each with its source.
05 · Later
Time series and plate scans, once a single image is trusted.
We are early stage. Stage 01 runs on example data and one public image.
Pipeline v0.1: a contrast gate, a focus gate, classical segmentation and size and shape rules. It runs on example data and on one public bright-field image. We publish no accuracy number until a held-out test prints one.
The pipeline could not measure what the rule needs: the image is blurry or empty, the organoid touches the frame, two objects overlap, or a value sits within 10% of its cut. It is an answer, not a failure, and it goes to a person.
The first rules are type-agnostic: size, circularity, solidity, edge and overlap. Floors per organoid type come with the first lab partners, each with its source.
Each upload is stored once under its sha256 and never rewritten. Every result names the pipeline version and its hash. Decisions and the audit log are append-only.
For organoid labs, screening teams and biobanks that want a recorded QC call, and investors who want to see the plan.
Render