Cancer Organoid Detection & Tracking

The problem

Cancer organoids are small, lab-grown versions of tumors used to test drugs. Researchers need to know when each organoid is ready for drug testing, and that meant checking every frame by hand.

My approach

Detect and track

A YOLOv5 model detects organoids in each frame, and DeepSORT follows each one over time as it grows.

Classify readiness

Each organoid is classified into one of 5 readiness stages, plus a class 0 for debris and non-viable organoids (nothing grew, or the cells had died). Class 0 was the most common class by far, so I handled the imbalance in two ways:

• Oversampling the rare readiness stages, so the model sees enough of them
• Weighting the loss by class frequency, so mistakes on rare stages cost more

Keep humans in the loop

Instead of checking every frame, researchers now only review what the model is unsure about:

• Predictions with high confidence pass automatically
• Uncertain predictions wait for human approval

This automated 89% of predictions, where annotation used to be 100% manual.

The data pipeline

I also architected and maintained the lab's 150 TB imaging data pipeline, working with clinical researchers at CHUV (Lausanne University Hospital).

What I learned

• Tie the model to a decision. "Is this organoid ready for drug testing?" is a question researchers actually ask, so the model's output fits straight into their work.
• Messy data deserves its own class. Debris and dead cells are real; giving them a class made the readiness stages cleaner.
• Automate the confident cases, not all of them. A human-in-the-loop pipeline saves most of the time while keeping experts in charge of the hard calls.