Brain tissue on a scan looks like a stormy sky: bright patches, dark hollows, blurred edges. The tumor hides among these patterns, and the doctor needs to outline its contour to plan surgery or radiation. One MRI mode highlights edema, another the tumor core, a third dead areas. But in a real hospital, some scans are lost or never taken: the patient moved, the machine broke, there's no time. The BMDS-Net program learns to work with what's available, like an experienced cook preparing dinner from an incomplete set of ingredients.
The secret is in two steps. First, the network gathers the available scans and decides on its own which one to trust more: if one mode is missing, it relies on the remaining ones, but not silently: it adds a correction for the missing data. At the same time, it learns to draw boundaries not at random but by the most reliable landmarks: where tissue changes color sharply, the contour is clear, and where the transition is gradual, the program marks it as a zone of uncertainty. The second step is honesty. Instead of a single answer, the network produces several variants, like several doctors looking at the same scan and slightly disagreeing. The spread of these variants is the measure of uncertainty.
The tests showed: when all scans are present, accuracy is maximal, but even if one mode is removed, the program barely loses its touch. And its confidence estimate was so accurate that it nearly matched an ensemble of three separate neural networks, although it trained at half the cost. For the doctor, this means: the program not only draws a contour but also hints where to double-check manually, like a navigator that doesn't stay silent about a bad road but warns about potholes in advance.
The accuracy of matching doctors' labeling is measured on the Dice scale: 1 is a perfect match, and 0.93 is like two people tracing the same shape and differing by only a couple of millimeters.