BraTS Brain Tumor Segmentation — Dice Score Explorer

Not medical advice. This page and its simulation are educational tools for understanding how segmentation AI is evaluated. They do not analyze real patient scans, do not diagnose any condition, and must never be used in place of assessment by a qualified radiologist, neurologist or oncologist. If you or someone you know has concerns about a brain tumor or any MRI finding, please consult a licensed medical professional.

Why Segmenting a Brain Tumor Is Harder Than It Sounds

To a non-specialist, "find the tumor on the scan" sounds like a single well-defined task. In practice, a glioma, the most common type of primary brain tumor studied in segmentation research, does not have one boundary; it has several, nested inside each other and each clinically meaningful in a different way. There is a bulk region of abnormal tissue and swelling visible mainly on T2-weighted and FLAIR MRI sequences, generally called the whole tumor. Inside that, there is a denser region including any necrotic or non-enhancing tissue, the tumor core. And within that, there is often a thin, contrast-enhancing rim that lights up specifically on a T1-weighted scan taken after a gadolinium contrast injection (T1ce), the enhancing tumor, frequently the most clinically significant sub-region because it can mark the most active, aggressively growing part of the mass. Radiologists trace these three overlapping regions by hand, slice by slice, through a 3D MRI volume, a slow, expertise-intensive process. Automating it well could meaningfully speed up treatment planning and monitoring, which is precisely why segmentation AI has become one of the most actively benchmarked problems in medical imaging, and why a rigorous, standardized way of scoring "how good" an automated segmentation is became essential.

Inside the BraTS Challenge: Multi-Modal MRI and Expert Ground Truth

The BraTS (Brain Tumor Segmentation) challenge, first proposed in work by Menze and colleagues and run annually since, is the field's reference benchmark for this problem. It assembles multi-institutional MRI scans of glioma patients contributed from hospitals and research centers around the world, each patient imaged with four standard, co-registered MRI sequences: T1-weighted (T1), contrast-enhanced T1-weighted (T1ce), T2-weighted (T2), and Fluid-Attenuated Inversion Recovery (FLAIR). Each sequence highlights different tissue properties, T1ce is best for spotting the enhancing rim, FLAIR is particularly sensitive to the swelling and infiltration that defines the whole tumor, so a good segmentation model has to learn to combine information across all four rather than rely on any single image. Alongside the raw scans, BraTS provides expert-annotated ground-truth labels: board-certified neuroradiologists (with results cross-checked by multiple experts) trace the whole tumor, tumor core, and enhancing tumor regions by hand on every training case. These hand-drawn labels are the "ground truth" that competing algorithms are measured against, exactly the role the fixed red blob plays in this simulation, a stand-in for that expert-drawn outline, deliberately irregular because real tumor boundaries are irregular, not the tidy circles a first attempt at automated segmentation often naively produces.

The Dice Similarity Coefficient, Explained

The core scoring formula is the Dice similarity coefficient (also called the Sørensen–Dice index): Dice = 2 × |A ∩ B| / (|A| + |B|), where A is the ground-truth region, B is the predicted region, |A| and |B| are their areas (or, for real 3D MRI volumes, their volumes), and |A∩B| is the size of their overlap. The factor of two compensates for the fact that the overlap region is otherwise counted only once while it is being compared against the sum of two full areas, so a perfect match, where A and B are identical, always produces a Dice score of exactly 1.0, and two regions that do not touch at all produce 0. Working through the arithmetic makes the formula's behavior intuitive. If two regions overlap over 80% of their combined footprint, Dice sits comfortably in the 0.85-0.95 range. If a predicted region is roughly the right size but shifted so only half of it lands on the true tumor, Dice drops sharply, often below 0.6, because both the false-positive area outside the true tumor and the false-negative area of true tumor left uncovered count against the score simultaneously. That dual sensitivity, penalizing both over-segmentation and under-segmentation in one number, is exactly why Dice became the default metric rather than something simpler like raw pixel accuracy, which can look deceptively high on a mostly-normal-tissue image even when the tumor itself is segmented poorly, since correctly labeling all the healthy tissue as "not tumor" dominates the accuracy score regardless of how the tumor region itself is handled. In this simulation, moving the predicted blob's offset sliders shifts it away from the fixed ground truth, directly reducing the overlap area |A∩B| while the individual areas |A| and |B| barely change, which is exactly why Dice falls fastest under pure positional drift. Shrinking or enlarging the size slider changes |B| itself, illustrating the second, independent way a segmentation can go wrong even when it is perfectly centered.

Why Real Leaderboards Also Report the Hausdorff Distance

Dice is an area-based (or volume-based) metric, and area-based metrics have a specific blind spot: a small, isolated spike of incorrectly predicted tumor tissue, sticking far out into otherwise healthy brain tissue, contributes only a tiny amount to the total mismatched area, so it barely moves the Dice score even though it could represent a serious, clinically dangerous localization error. This is precisely the kind of failure a boundary-distance metric is built to catch. The Hausdorff distance measures worst-case boundary disagreement rather than overlap area. Informally: for every point on one region's boundary, find the distance to the nearest point on the other region's boundary; take the largest such distance found (this is the "directed" Hausdorff distance in one direction); do the same calculation in the other direction; the true Hausdorff distance is the larger of those two directed values. A large Hausdorff distance means that somewhere along the boundary, the two outlines diverge badly, even if the overall Dice score still looks reasonable. Real BraTS-style leaderboards typically report a 95th-percentile variant of this distance rather than the strict maximum, which trims out single-pixel noise while still catching genuine large boundary errors, but the underlying idea is the same worst-case-mismatch logic implemented here in simplified, clearly-labeled form. This simulation's "boundary mismatch" indicator is a deliberately simplified, illustrative approximation of that same idea, computed by sampling points around each blob's boundary and finding the largest nearest-neighbor gap between the two point sets. Watch what happens when you push the jaggedness slider up: the Dice score often barely moves if the overall size and position stay similar, but the boundary-mismatch number can jump noticeably, because a jagged edge creates exactly the kind of localized spikes and gaps that an overlap-only metric tends to understate.

Reading the Score: Excellent, Clinically Usable, or Needs Review

Raw Dice numbers only become meaningful once they are anchored to some practical reference point, which is why segmentation-challenge leaderboards and published clinical-validation studies commonly group results into qualitative bands. This simulation uses three such bands. A Dice score above 0.85 is treated as excellent overlap, comparable to the agreement seen between two different expert human annotators tracing the same tumor independently, sometimes called inter-rater agreement, a useful ceiling because an AI model rarely needs to exceed how consistently two trained radiologists agree with each other. A Dice score between 0.70 and 0.85 is labeled clinically usable, generally good enough to assist workflows like rough treatment-planning drafts or longitudinal tumor-volume tracking, though still meriting a clinician's review rather than fully automated reliance. A Dice score below 0.70 is flagged as poor — needs review, the range where a segmentation is likely missing significant tumor tissue, including significant healthy tissue by mistake, or both, and should not be trusted without correction. These bands are not arbitrary; they reflect thresholds that repeatedly show up across published segmentation-validation literature as rough dividing lines between "matches expert-level agreement," "usable with oversight," and "not yet reliable enough." They are, however, general-purpose guidance rather than a fixed clinical standard, actual acceptability thresholds vary by institution, by tumor sub-region (enhancing-tumor Dice is typically harder to achieve than whole-tumor Dice because that region is smaller and its boundary more subtle), and by the specific downstream use of the segmentation.

What This Simulation Simplifies, and Why That's Fine for Learning

A few honest simplifications are worth naming. Real BraTS segmentation happens in three dimensions across a full MRI volume, not on a single flat 2D shape; this simulation works in 2D purely so the overlap and boundary-mismatch behavior stay directly visible and intuitive rather than requiring a volumetric renderer to appreciate. Real ground-truth tumor boundaries come from actual expert annotation on actual patient scans, not from a fixed mathematical noise function; the irregular red blob here is a stand-in shaped to feel anatomically plausible, not an image of any real tumor. And the boundary-mismatch indicator is explicitly a simplified proxy for the Hausdorff distance rather than the exact 95th-percentile calculation used in published BraTS results. None of that undermines the point of the exercise. The relationship between position, size, and jaggedness on one side, and Dice score and boundary mismatch on the other, behaves the same way here as it does in real volumetric tumor segmentation, because both are built on the same overlap and boundary-distance mathematics. Dragging the sliders and watching the numbers respond is, in miniature, the same feedback loop a machine-learning engineer sees when they retrain a segmentation model and re-run it against the BraTS validation set.

Frequently asked questions

Try it live

Everything above runs in your browser — open BraTS Brain Tumor Segmentation — Dice Score Explorer and change the parameters while it is running. Nothing is installed, nothing is uploaded, the whole model lives in one tab.