A mole on your skin looks like a mosaic of colored pieces. Each piece is a small patch of the image where color and texture barely change. Such pieces are called superpixels. If you connect them with lines, you get a network that looks like a subway map: stations are the pieces, and lines are the connections between them. The weight of each line shows how strongly two pieces are related. Scientists tried two ways to assign these weights: either use a ready-made formula, such as the distance between colors, or let the computer figure out the weights itself so that the network better distinguishes melanoma from a regular mole.
They built not just one network but several, with different numbers of pieces — from 20 to 100. This creates an ensemble: one network sees large details, another sees small ones. In addition, weak connections were pruned, leaving only the most important ones, and they checked how this affected the result. It turned out that the best results came from networks with computer-tuned weights and signals that describe the texture of a piece, not just its color. Accuracy on test images reached 99%, meaning that out of a hundred moles, the computer made a mistake in only one.
Such an approach could be built into software for dermatologists: a doctor photographs a mole, and the system immediately suggests whether the patient should be sent for a biopsy. This is faster and cheaper than waiting for a specialist consultation.
Superpixels are like a puzzle where the pieces automatically adapt to the boundaries of objects in the photo. The more pieces, the finer the details the computer can see.