Case study · India
Early failure detection for cell therapy manufacturing
Daily microscopy images, measured cell by cell, feed a model that predicts which 120-day culture batches are likely to fail, days or weeks earlier than before.
- Daily microscopy images
- Cellpose3 segmentation
- Per-cell morphometry
- Gradient boosting + fine-tuned CNN
- Early warning for scientists
The problem
A cell therapy manufacturer grows cell cultures in batches that run for about 120 days. Around 40% of them fail, and the failure is often found late, after weeks of media, incubator time and skilled work have gone into a batch that was not going to succeed. Scientists image every culture under the microscope each day, but spotting a slow change in cell shape across months of images by eye is hard to do consistently.
What we built
A decision-support tool that reads the daily microscopy images and predicts how each batch is likely to end, early enough to act on. It flags batches that look like they are heading for failure and shows the measurements behind each flag.
It is decision support only. The model does not stop, continue or change any batch. Scientists see the prediction and the evidence, and keep every call.
How it works
- Finding the cells. Each image is segmented with Cellpose3, which outlines every individual cell.
- Measuring them. For each image we compute morphometric features: cell count, size, shape, how the population is distributed, and how all of these change from one day to the next.
- Two models, combined. A gradient boosting model works on the measured features, which keeps its reasoning inspectable. A convolutional neural network, fine-tuned on the images themselves, picks up visual patterns the measured features miss. A hybrid of the two produces the prediction for each batch.
- Earlier is the point. Predictions use only the images available so far, so a batch can be flagged days or weeks before its failure would otherwise be noticed.
Outcome
The aim is to move the moment a failing batch is spotted from late in the run to days or weeks earlier, so the lab can stop it and put the capacity to better use. Scientists get an early, evidence-backed signal on every batch, and every call stays with them.
Stack
| Part | Tools |
|---|---|
| Segmentation | Cellpose3 |
| Features | Per-cell morphometry, tracked day by day |
| Models | Gradient boosting on cell features, fine-tuned CNN on images |
| Language | Python, PyTorch |
Have a problem like this one?
We reply within one business day. If AI isn't the right fix, we'll say so.