This simulation models how automated brain-tumor segmentation AI is scored against a human-drawn "ground truth" region, the core evaluation method used in the real BraTS (Brain Tumor Segmentation) challenge. A fixed irregular blob represents a radiologist-annotated tumor outline on an MRI slice; a second, adjustable blob represents an AI model's predicted segmentation.
The Dice similarity coefficient is computed as 2×|A∩B| / (|A|+|B|), where A is the ground-truth region and B is the predicted region — exactly the formula used to rank real segmentation-challenge submissions. A simplified, clearly-labeled boundary-mismatch indicator, inspired by the Hausdorff distance used alongside Dice in real evaluations, estimates how far the two outlines drift apart at their worst-matching points.
Horizontal offset slider, vertical offset slider, size/radius slider, and a jaggedness/noise slider, all applied to the predicted (blue) segmentation blob. The ground-truth (red) tumor outline stays fixed.
Did you know that in the real BraTS challenge, teams are scored on three overlapping tumor sub-regions at once, whole tumor, tumor core, and enhancing tumor, each with its own Dice score, because a model can localize a tumor well overall while still missing its most clinically important enhancing core?
This simulation models how automated brain-tumor segmentation AI is scored against a human-drawn "ground truth" region, the core evaluation method used in the real BraTS (Brain Tumor Segmentation) challenge. A fixed irregular blob represents a radiologist-annotated tumor outline; a second, adjustable blob represents an AI model's predicted segmentation, scored live with the Dice similarity coefficient and an illustrative boundary-mismatch indicator.
This simulation models how automated brain-tumor segmentation AI is scored against a human-drawn "ground truth" region, the core evaluation method used in the real BraTS (Brain Tumor Segmentation) challenge. A fixed irregular blob represents a radiologist-annotated tumor outline; a second, adjustable blob represents an AI model's predicted segmentation, scored live with the Dice similarity coefficient and an illustrative boundary-mismatch indicator.
Drag the horizontal and vertical offset sliders to shift the predicted (cyan, dashed) segmentation away from the fixed ground-truth (red) tumor outline, adjust the size slider to make it too small or too large, and increase jaggedness to make its edge noisier and less anatomically plausible. Watch the Dice score, the overlap quality band, and the boundary-mismatch indicator update live as the two outlines drift apart or converge.
Did you know that in the real BraTS challenge, teams are scored on three overlapping tumor sub-regions at once, whole tumor, tumor core, and enhancing tumor, each with its own Dice score, because a model can localize a tumor well overall while still missing its most clinically important enhancing core?