The problem — Global pixel overlap scores like Dice are insensitive to the errors that actually matter in live-cell and time-lapse microscopy: an object that is present but poorly delineated, an object that is merged with its neighbor, or an object that vanishes entirely from the predicted masks. A workflow that counts, tracks, or lineages cells needs per-instance correctness — and SEG provides it.
What it is / how it works — The SEG measure, defined in the 2017 Cell Tracking Challenge paper by Ulman et al., computes the mean Jaccard similarity (IoU) over all reference cell instances. For each reference object, the algorithm finds the predicted mask that maximally overlaps it; if the overlap exceeds a matching threshold (typically IoU > 0.5), the Jaccard value for that pair is recorded. The final SEG score is the average over all reference objects, including unmatched ones (which score 0), so missed cells and false-positive regions both drive the score down. This makes SEG sensitive to both detection errors and delineation quality simultaneously.
SEG is reported alongside TRA (tracking accuracy) in the Cell Tracking Challenge, which has served as the standard benchmark for cell segmentation and tracking algorithms across 2D and 3D fluorescence microscopy modalities since 2013. The 10-year retrospective (Maška et al., 2023) confirmed it as the go-to instance-level benchmark for live-cell workflows.
Where it breaks — SEG is designed for fluorescence time-lapse data with annotated cell instances. It requires a ground-truth segmentation at the instance level — which is expensive to produce and, as Ground Truth Is a Design Problem argues, always reflects the annotator's decisions about ambiguous boundaries. In workflows without labeled instances (brightfield, label-free, spatial omics), SEG cannot be computed directly. And like IoU at a fixed threshold, it can reward a model that matches most cells while systematically failing on a biologically important rare subpopulation.
The Fovea stance: SEG is one of the most honest segmentation metrics available for live-cell and tracking workflows because it evaluates each object individually rather than averaging pixels. It belongs alongside Panoptic Quality and object-level F1 in the evaluation toolkit for any workflow-level segmentation audit — not as a replacement for domain validation, but as a principled first screen.