StarDist

Two nuclei touching at an oblique angle share almost the same axis-aligned bounding box, so a detector that localizes boxes and refines them suppresses one of the pair as a duplicate. A network that instead labels pixels as nucleus or background and groups connected components makes the opposite mistake: a few misclassified boundary pixels fuse the two into one object. Both failures concentrate in the densely packed images biologists most want counted. StarDist sidesteps the choice with a shape accurate enough to need no refinement — a star-convex polygon, given by the radial distances from an interior pixel to the object boundary along fixed directions.

Uwe Schmidt, Martin Weigert, Coleman Broaddus and Gene Myers introduced the method at MICCAI 2018 from the Max Planck Institute of Molecular Cell Biology and Genetics and the Center for Systems Biology Dresden, with Myers also at TU Dresden. The network predicts, for each pixel, an object probability and the full distance vector of the object containing it; greedy non-maximum suppression over that redundant pool keeps one polygon per object. Object probability is the normalized Euclidean distance to the nearest background pixel, so suppression prefers the polygons proposed from pixels near a nucleus centre — the ones that describe the shape best.

The package has since grown two extensions: star-convex polyhedra for volumes (WACV 2020, with collaborators at ENS de Lyon and CNRS), and an instance-classification head for histopathology (ISBIC 2022, at EPFL). With Cellpose, it is one of the two segmentation engines most bioimage pipelines are built around.

#Key Features

• Shape model that skips refinement: Radial distances along a fixed ray set describe roundish nuclei closely enough that the polygon is the final answer, not a proposal to be masked afterwards.
• Pretrained generalist checkpoints: 2D_versatile_fluo for fluorescent nuclear markers and 2D_versatile_he for brightfield H&E load by name from a hash-pinned registry and segment new images with no training or annotation.
• Anisotropy-aware 3D: Star-convex polyhedra sample rays on a spherical Fibonacci lattice and rescale them for the squeezed nuclei that anisotropic voxel spacing produces in confocal stacks.
• Segmentation plus classification: An extra semantic head assigns each instance a discrete class by aggregating per-pixel class probabilities over that instance's own pixels.
• Distribution through GUIs: The pretrained models run from Fiji, QuPath, napari, Icy and KNIME, reaching labs that write no Python.

#Technical Details

The 2D model is a U-Net with an added 3x3 convolutional layer of 128 channels feeding two heads: a single-channel sigmoid for object probability and an n-channel linear head for distances, with n = 32 rays. Training minimizes binary cross-entropy on the probability map and an object-probability-weighted mean absolute error on the distances, so background pixels contribute nothing and central ones count most. On a cleaned 497-image fluorescence subset of the 2018 Data Science Bowl it reached average precision 0.864 at IoU 0.5, against 0.832 for Mask R-CNN, 0.806 for a three-class U-Net and 0.674 for a two-class U-Net, with roughly an order of magnitude fewer parameters than Mask R-CNN's ~45 million.

Extending 32 planar rays naively to 3D would demand 1,024 directions; the Fibonacci-lattice construction is faithful with 64, and suppression replaces exact polyhedra intersections with successively tighter bounds. The histopathology variant uses 64 rays and a depth-4 U-Net trained on 4,981 CoNIC patches from the Lizard dataset across six nucleus classes, and took first place on both the preliminary and final CoNIC 2022 leaderboards for segmentation and classification. The released checkpoints come from modest public corpora — the Data Science Bowl subset for fluorescence, MoNuSeg 2018 plus the TNBC dataset for H&E — and ship from a separate repository under the code's BSD-3-Clause licence. Only a demo checkpoint is registered for 3D, so volumetric work starts from user-labelled data.

#Applications

The fluorescence checkpoint is a default first pass for counting DAPI- or Hoechst-stained nuclei in screening plates, time-lapses and tissue sections; the H&E checkpoint does the same for digitized pathology slides, where nucleus counts and morphology feed cell-type and grading analyses. Because instances arrive as polygons with plausible completions at image borders, the output suits tracking pipelines that need several segmentation hypotheses per frame. Labs with unusual staining retrain on a handful of annotated crops rather than starting over.

#Impact

StarDist became a component of bioimage analysis rather than a standalone tool — the segmentation stage inside larger workflows, and the baseline newer segmentation papers report against. Its limits are structural and openly stated: objects must be roughly star-convex, so elongated or lobed shapes need many more rays; the parametric representation loses accuracy at high IoU thresholds relative to per-pixel segmentation; and the pretrained checkpoints cover only 2D in two staining regimes. The dependence on TensorFlow and a compiled C++ extension makes installation the most common practical obstacle.