Detecting tumor indicators from cell images

The problem

Examining slides under a microscope is slow and tiring: a doctor examines dozens of slides a day. Fatigue lowers accuracy, while early cases are exactly the ones that need the most focus.

How it works

1. PreparationNormalizing stain colors and splitting the slide into tiles.
2. TrainingTraining a neural network on tiles labeled with the correct class.
3. ClassificationGiving each tile a confidence score.
4. AssemblyMerging the tile results into a heat map and prioritizing the whole slide.

Project features

• Prioritization rather than a final verdict: the tool assists, it doesn't replace.
• A heat map shows which regions the model relied on, so the doctor can review its decision.
• Performance is measured by sensitivity, not overall accuracy, because missing a case is worse than a false alarm.
• Making up for few labeled samples with data augmentation.

Expected challenges

These are the main obstacles to expect during the build, laid out clearly so you know them before you start.

• Labeled medical data is rare and access-restricted — the first obstacle.
• Stain differences between labs make the model fail unless normalization is done.
• Class imbalance: healthy cases far outnumber the rest, so the model tends to call everything healthy.
• Interpretability is a medical requirement; a black-box model isn't accepted in this field.

The consultation gives you a solution for each one.