Flood-Filling Network

A Flood-Filling Network works much like the paint-bucket tool in a drawing program. Click one point on an image and the paint bucket instantly fills the whole connected region of the same color. This technique does something similar: starting from a single pixel, it keeps checking whether each neighboring pixel belongs to the same object, and grows the region outward pixel by pixel.

This approach is used in brain mapping. When a brain is sliced into millions of cross-sections thinner than a human hair and photographed, each image shows countless neuron fibers tangled together. Manually marking which fiber belongs to which neuron across all those images would take an enormous amount of time. This technique automates that work by starting from one pixel, judging whether neighboring pixels belong to the same neuron, and expanding the region accordingly. Stitching together millions of these cross-sections this way reveals the full 3D shape of each individual neuron and how it connects to others.

This technique is one of the core methods a Google research team has used to reconstruct 3D neural structures from flat electron microscope images. It later evolved into a newer system trained with a mix of synthetic data, which was used to map the fruit fly brain.

The article explains that a "convolutional neural network-based Flood-Filling Network" is the core technology for reconstructing 3D neural structures from flat electron microscope images. It's worth remembering that this term doesn't refer to the finished brain map itself, but to the reconstruction algorithm used to build that map.