Loading…
Spot detection and quantification in 5D fluorescence microscopy. Pretrained 2D and 3D U-Nets segment foci, then Gaussian fitting measures each one.
No providers recorded yet. Browse all providers
A fluorescent focus in a confocal z-stack spans several planes. A detector run slice by slice reports it three times; a biologist scrolling the stack by hand can just as easily miss it once. Both errors come from discarding the third dimension before deciding what counts as one object. SpotMAX treats the whole acquisition as the unit of analysis instead: z-slices, time points and every fluorescence channel enter the same decision, and each candidate spot is scored inside the segmentation mask of the cell containing it.
SpotMAX was developed at the Institute of Functional Epigenetics of Helmholtz Munich with the Cell Biology and Biophysics Unit at EMBL Heidelberg. Its pretrained component, SpotMAX AI, is a pair of U-Net semantic-segmentation networks — one 2D, one 3D — shipped as fixed checkpoints the Python package downloads the first time a model is instantiated. New acquisitions are mapped onto the training resolution by rescaling, not by refitting.
The networks sit inside a four-stage pipeline: optional segmentation of a reference channel such as the mitochondrial network, segmentation of the spot channel, detection by local-maximum search, and quantification by 3D Gaussian fitting. Each stage accepts external input, so the U-Nets can be swapped for thresholding, for Spotiflow, or for any BioImage.IO zoo entry — where the 2D checkpoint is itself published as SpotMAX-AI. That model, not the wrappers, is what is catalogued here.
Both networks are encoder-decoder U-Nets in PyTorch, built from batch normalisation, convolution, ReLU and max-pooling blocks with encoder feature maps concatenated into the matching decoder stage and a final 1x1 convolution producing background and foreground classes. The 2D model is custom; the 3D model is built on pytorch-3dunet with feature maps of 32, 64, 128 and 256. Training targets were binary spot masks made by placing spheroids of Abbe diffraction-limit radii at published spot centres from Cell-ACDC-analysed smFISH and mitochondrial DNA datasets, plus deepBlink's 2D "smfish" and "suntag" sets, then hand-corrected by an independent expert — 4,956 training, 1,471 validation and 1,586 test z-slices. Preprocessing rescales to 73 nm/pixel, strips hot pixels, and MinMax-scales to (-1, 1). The 2D model trained for 100 epochs at batch size 16 on 400x400 crops with an inverse-Dice loss; the 3D model on (30, 250, 250) patches for 15,000 steps, both with SGD on a single 32 GB V100.
The benchmark ground truth covers 369 z-stack volumes with 20,692 annotated spots from 2,894 cells, spanning smFISH in S. cerevisiae, mitochondrial DNA nucleoids, and COSA-1 crossover foci in C. elegans. Because part of that data trained the U-Nets, the comparisons against RS-FISH, Spotiflow and human annotators use SpotMAX's thresholding path rather than SpotMAX AI. On crossover counting, where each of the six C. elegans homolog pairs receives exactly one crossover, SpotMAX averaged 6.01 +/- 0.16 foci per nucleus against a human annotator's 5.98 +/- 0.24.
The paper works four biological questions end to end: meiotic crossover counting in C. elegans, multigenerational mitochondrial DNA dynamics across a nutrient shift in budding yeast, telomere length versus cell size in mouse hematopoietic stem cells, and Raptor-depletion effects on nucleus and nucleolus size. In each, spot counts alone were insufficient — the result depended on a quantified feature (integrated intensity for telomere length, network volume for mitochondria) attributed to a tracked single cell. It therefore suits wet-lab groups doing smFISH, DNA-FISH or nucleoid imaging who need per-cell numbers out of a time-lapse, not an overlay.
SpotMAX's most durable contribution may be the annotated 3D dataset released alongside it, the first public ground truth of its kind for 3D spot detection and deposited in the BioImage Archive for reuse. The framework is positioned as glue rather than replacement: it is a plugin to the same lab's Cell-ACDC and hosts competing detectors as first-class backends. Two limitations are worth stating plainly: the headline benchmark does not measure SpotMAX AI, since the U-Nets saw part of the evaluation data during training, and the checkpoints target diffraction-limited globular foci at one pixel size, so structures far from that regime fall outside what rescaling can compensate for.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.