The problem — Dice is ubiquitous on segmentation leaderboards, and a high Dice number is routinely treated as evidence that a model is ready for a microscopy workflow. It is not. Dice reports semantic overlap — the fraction of correctly classified pixels — not instance correctness: whether the right number of objects are present, whole, and separate.
What it is / how it works — The Dice Similarity Coefficient computes twice the size of the intersection between a predicted and reference mask, divided by the total size of both masks. Equivalently, DSC = 2 × TP / (2 × TP + FP + FN) at the pixel level. It penalizes false positives and false negatives equally and is naturally bounded between 0 (no overlap) and 1 (perfect agreement). Because it operates on the full pixel map, large foreground regions dominate the score; small or rare structures contribute little even when they carry the biology.
In practice Dice is a fast, differentiable sanity check that captures gross segmentation quality and is a reasonable training loss (Dice loss) for semantic tasks. Metrics Reloaded classifies it under semantic segmentation metrics and notes that it is only appropriate when the task is truly semantic — i.e., when individual object identity does not matter.
Where it breaks — In Cell Painting and most live-cell workflows, biology is counted per object: cells, nuclei, organoids. A model that merges two adjacent cells into one mask loses no more than a few pixels of perimeter, so Dice stays high while the instance count is wrong. The companion pitfalls paper catalogs how this happens systematically in dense, confluent fields — precisely the images most common in high-content screening. In digital pathology and spatial omics, boundary errors that are invisible to Dice misassign transcripts or organelle memberships to neighboring cells, corrupting the readout.
The Fovea stance: Dice is a useful first filter and a practical training objective, but it answers the wrong question for any workflow where objects are counted, tracked, or individually profiled. For those tasks, Evaluating Segmentation Models in Real Microscopy Workflows calls for instance-level metrics such as IoU-at-threshold object F1, Panoptic Quality, or the Cell Tracking Challenge SEG measure.