Spotiflow

Imaging-based spatial transcriptomics encodes transcript identity combinatorially across multiple imaging rounds, so a single missed or spurious punctum in one channel propagates into a misassigned gene. Detecting those puncta has for decades been handled by Laplacian-of-Gaussian or radial-symmetry filters, which work well on clean data and require a human to choose an intensity threshold per channel, per round, and often per field of view. In a gigapixel MERFISH or HybISS experiment that tuning is both the bottleneck and the main source of irreproducibility.

Spotiflow, from the Weigert and La Manno labs at EPFL with collaborators at EMBL Heidelberg and TU Dresden, replaces the filter with a convolutional network trained on two complementary targets. The first is a multi-scale hierarchy of Gaussian heatmaps that localize spot centers coarsely. The second — the method's namesake — is stereographic flow: instead of regressing the raw vector pointing from each pixel to its nearest spot, the network regresses that vector's inverse stereographic projection onto a sphere. Offsets from pixels far from any spot all map toward a single pole, which removes the indeterminate-target problem that makes plain offset regression unstable in sparse images.

The design generalizes cleanly to arbitrary dimension, so the 3D version predicts volumetrically rather than detecting per plane and blending. On synthetic volumetric data the native 3D model reaches an F1 of 0.882 while 2D models with Z-blending collapse to 0.102 and 0.311.

#Key Features

• No threshold to tune: A single trained model produces detections directly, eliminating the channel- and round-specific threshold search that classical iST pipelines require.
• Subpixel localization: Peaks in the highest-resolution heatmap give preliminary coordinates, which the inverted stereographic flow refines below the pixel grid.
• Native 3D prediction: Volumetric extensions of both targets handle anisotropic voxels and depth-dependent illumination without a plane-wise blending step.
• Pretrained checkpoints out of the box: Named models — a general model trained across all benchmark datasets, plus HybISS, live-cell, 3D smFISH and synthetic variants — load in one call from a separate BSD-3 weights repository and run on new images with no retraining.
• Cheap adaptation: Fine-tuning a synthetic-pretrained model on four images of a new live-cell dataset reached F1 0.738 against 0.174 for training from scratch; three DNA FISH images lifted the general model from 0.77 to 0.92.
• Autofluorescence discrimination: Trained against a dedicated autofluorescence round, the model halves the number of lipofuscin-derived false spots in adult mouse brain sections.

#Technical Details

The backbone is a U-Net predicting both targets from network decoder feature maps at several resolutions, which jointly contribute to the training loss and improve convergence when spots are sparse. On a realistic synthetic benchmark including autofluorescence and optical aberrations, Spotiflow reached F1 0.929 against 0.915 for the deep learning method deepBlink and 0.758–0.836 for classical detectors. On real annotated data the margin over threshold-based methods was largest where background is worst: F1 0.796 on HybISS and 0.861 on MERFISH, against 0.531 and 0.790 for a Laplacian-of-Gaussian detector. A single general model trained on all datasets matched per-dataset models almost exactly.

Efficiency was engineered alongside accuracy. Prediction on a 32k × 32k image takes 80 s versus roughly 1000 s for the LoG pipeline it replaces, with over three times lower peak memory; a 159 GB EASI-FISH volume ran in under an hour on one GPU and under ten minutes across eight A100s. Training a 2D model takes roughly an hour on a single RTX 4090. Inference parallelizes near-linearly across GPUs and reads Zarr data directly from cloud object storage.

#Applications

Spotiflow is the transcript-detection step in imaging-based spatial transcriptomics — MERFISH, seqFISH, HybISS, smFISH — and it slots into existing pipelines through a Starfish integration, a napari plugin, a QuPath extension, and a TrackMate module. Substituting it for the default detector in an end-to-end HybISS experiment on developing mouse brain produced gene expression maps consistent with reference in situ hybridization atlases while removing per-channel threshold choices entirely. Beyond transcriptomics it handles live-cell tracking of telomeres and TERRA foci, 3D lipid droplet tracking in label-free movies, and even non-spot objects such as microbial colonies on agar plates.

#Impact

Spot detection was one of the last steps in the spatial transcriptomics stack still dominated by hand-tuned signal processing, and Spotiflow is the entry that changed the default. Its practical significance is less the accuracy margin than the removal of a manual step that scaled with experiment size, which is what makes organism-scale iST data collection tractable. Published in Nature Methods, distributed BSD-3 on PyPI and conda-forge with a CLI, Docker image, and plugins for four analysis platforms, it has been broadly adopted as a drop-in replacement in existing workflows. The main dependency is annotation quality: subpixel accuracy is bounded by how precisely the training spots were marked, and imaging conditions far from the pretraining distribution still call for a handful of fresh annotations.