AI Toxicity | AiderBio

From image upload to toxicity call

The software analyzes live/dead fluorescence and time-lapse brightfield images, learns how they change over time, and predicts drug response. The same criteria apply to every image, and hundreds of images are processed in one run.

AI well readout — Automatic well readout. Green: readable, red: optically blocked (bubble or unfilled media suspected)

Well detection and quality check

Finds Port Pocket microwells at any magnification and excludes wells blocked by bubbles or unfilled media.

Segmentation

Segments organoids in brightfield and fluorescence images and measures area, shape and live/dead signal per well.

Early toxicity prediction

Uses early time points to predict the endpoint toxicity grade (T0, T1, T2) and viability.

Labeling, retraining and reports

Corrections become training data. Results export as PDF and Excel with a full audit trail.

Validated on AiderBio images

We tested well detection and readout on 29 Port Pocket images containing 229 microwells. Readout accuracy was measured on culture conditions that were not used to set the model.

0.996Well detection recall (228/229)

1.000Well detection precision

0.979Readout accuracy, held-out conditions (142 wells)

0.9996Readout AUC

Reference labels by one AiderBio reader; confirmation by two or more readers is planned.

Model tracks

Track 1 · Anomaly detection

A transformer-based time-series model reads day 1–7 brightfield images (shape, size, density) and flags when 3D growth becomes abnormal.

Normal vs. abnormal ROC, internal prototype AUC 0.87

Track 2 · Viability prediction

A U-Net model learns from Calcein-AM time-lapse images paired with MTT viability and predicts viability from images alone.

Predicted vs actual — Predicted vs. measured viability, internal prototype R² 0.85

Our AI turns organoid images into consistent toxicity calls