Count by dividing
CNN cell quantification · IEEE Access
A neural network takes a fixed-size input, but microscope images come in every size and can hold hundreds of cells. The trick: cut the image into tiles, have a small CNN count each, and add them up. Simple — but tile size matters a lot.
FIG. 1 — Tile, count, sum · interactive
Tile, count, sum
Each cell in the grid shows the CNN’s predicted count for that tile. Change the tile size and watch the total drift away from the true count — then find the size where it’s most accurate.
Count error vs tile size — best near the ~100 px training size
FIG. 2 — The Goldilocks tile
The Goldilocks tile
Make the tiles too big and each one packs in many cells; a regressor trained on modest counts saturates and undercounts the crowded tiles. Make them too small and cells start landing on tile borders, getting counted in two neighbors at once — an overcount. The sweet spot is right around the ~100 px tiles the model was trained on, which is exactly why the training tile size is a real design decision, not an afterthought. The error curve makes the tradeoff visible: it dips to near zero in the middle and climbs at both ends.
The payoff of tile-and-sum: one small, fast model handles a 200×200 image or a 2000×2000 one with no retraining — it just runs on more tiles. That keeps CellQuant light enough to serve from a $12/month box and quick enough to feel instant in the browser.
Working on cell counting, digital pathology, or point-of-care imaging? Get in touch.
Email me