Deep learning cell segmentation, nucleus detection, and automated quantification — running natively inside ImageJ and Fiji. Priced for the independent investigator. Accurate enough for the core facility.
The annotation bottleneck in quantitative cell biology is not a data problem — it is a labour problem. High-content screening instruments generate tens of thousands of fields per experiment. The cells in those images do not label themselves. Manual correction at that scale is not a method; it is a funding liability. FluorocellAI was built to eliminate that liability without asking the investigator to become a machine learning engineer.
TIFF, CZI, LIF, ND2, ICS/IDS. Direct OMERO server connection. No format conversion at any stage.
Illumination correction, background subtraction, flat-field normalization, multi-channel registration.
Transformer encoder + CNN decoder ensemble. Nuclear detection, boundary delineation, cell-cycle phase classification.
Per-field confidence scoring, outlier flagging, manual correction via native ImageJ ROI Manager.
CSV, HDF5, ZARR, label TIFF, ROI Manager. Direct to CellProfiler, R, Python, or custom pipelines.
Installs in three clicks via the Fiji update site mechanism. No separate application. No workflow disruption.
High-content fluorescence microscopy platforms — Operetta, CQ1, IN Cell Analyzer — routinely generate datasets of tens of thousands of fields per experiment, each containing hundreds to thousands of individual cells. The acquisition side of this problem has been solved. The downstream side has not. Identifying individual cells within those images remains the primary constraint on throughput in quantitative cell biology.
Classical segmentation approaches fail systematically on the conditions most relevant to high-content work. Otsu and multi-level thresholding cannot resolve boundaries between adjacent cells in confluent monolayers. Watershed applied to nuclear DAPI staining produces systematic errors in S and G2 phase cells where nuclear morphology departs from the ellipsoidal model the algorithm assumes. These are not edge cases — they are the normal operating conditions of the assays that generate the most biologically important data.
FluorocellAI replaces this bottleneck with a purpose-trained deep learning inference pipeline. The underlying model was trained on a curated dataset of >2.8 million manually annotated cells across twelve cell lines, six imaging platforms, and a broad range of staining conditions — including confluent monolayers, three-dimensional organoids, and primary cultures with irregular morphologies that generalist models handle poorly.
The system integrates within ImageJ and Fiji — the environments the cell biology community has standardized around — preserving accumulated analytical infrastructure: macros, measurement pipelines, plugin ecosystems. FluorocellAI adds capability without extracting ownership of the workflow. The investigator retains full control.
Precise delineation of individual nuclei in densely packed monolayers and three-dimensional organoid sections. Handles touching and overlapping nuclei through learned instance separation — not post-hoc watershed. Validated across HeLa, U2OS, A549, MCF-7, primary hepatocytes, iPSC-derived neurons, and murine embryonic fibroblasts.
Cytoplasmic membrane segmentation for cells expressing membrane-targeted fluorescent proteins or stained with cytoplasmic dyes. Accurately resolves boundaries in confluent monolayers where adjacent membranes are in direct contact. Produces per-cell area, perimeter, eccentricity, solidity, and texture features compatible with CellProfiler pipelines.
Spectral co-localization and intensity quantification across up to eight simultaneous fluorescence channels. Automated per-cell Pearson's r and Manders' M1/M2 coefficients. Population-heterogeneity analysis at the single-cell level — not the image level — which bulk statistics systematically obscure.
Automated phase assignment using integrated nuclear DAPI intensity combined with EdU or BrdU incorporation signals where available. In the absence of S-phase markers, morphology and integrated intensity distributions construct the G1/G2 diploid-tetraploid ratio. Phase assignments exported as per-cell categorical variables for population-level analysis.
Plate-level batch execution for high-content screening datasets. Automated per-well QC: focus scoring, cell density assessment, saturation flagging, edge-effect detection. Aggregated phenotype statistics export in formats compatible with KNIME, CellProfiler Analyst, and R/Bioconductor. Z-score and SSMD calculations for hit identification pipelines.
Extension of the segmentation pipeline to z-stack volumes acquired by confocal, spinning-disk, light-sheet, or two-photon modalities. Produces volumetric cell meshes with 3D morphometric features: volume, surface area, sphericity, principal axes, inter-cell contact area. Compatible with Imaris, Napari, and the Fiji 3D Viewer.
Frame-to-frame tracking using nearest-neighbor with learned appearance features for disambiguation at track crossings. Produces cell lineage trees, migration trajectories, and per-cell temporal feature vectors. Validated at frame intervals from 2 minutes to 2 hours across widefield and confocal timelapse protocols.
Full programmatic access via the ImageJ Macro Language, BeanShell, Groovy, and Python through the SciJava framework. Enables embedding within custom pipelines, unattended HPC batch processing, and LIMS integration. REST API available for cloud inference deployments.
All F1 scores at IoU threshold 0.5. Cellpose 2.0 and StarDist results reproduced from published benchmarks using default parameters. FluorocellAI results use the pretrained general model without dataset-specific fine-tuning.
FluorocellAI installs as a standard Fiji plugin via the update site mechanism — the same mechanism used by every other plugin in the Fiji ecosystem. All segmentation outputs are written as native ImageJ ROI objects into the ROI Manager, making them immediately accessible to any downstream ImageJ operation without format conversion, data export, or workflow interruption.
The full ImageJ Macro Language API is exposed, enabling FluorocellAI inference to be scripted into automated pipelines running unattended on large image datasets overnight on institutional compute infrastructure.
Connects directly to the OMERO.server API. Segmentation results stored back as ROI and measurement table annotations, preserving full data provenance within your image management system. Compatible with OMERO 5.6+.
Export 16-bit label TIFF images as IdentifyPrimaryObjects input. Existing CellProfiler pipelines require no modification to downstream steps — only the segmentation step improves.
ZARR and HDF5 label arrays readable by NumPy and scikit-image. Napari-compatible label layers for interactive correction. A companion yashara.r package parses FluorocellAI HDF5 output into tidy data frames.
Tell us your cell type, imaging modality, and what you are trying to solve. We respond within two business days with a substantive reply — not a boilerplate acknowledgement.
For urgent or highly technical questions, write directly to the FluorocellAI team. This address is monitored by the scientists who built the segmentation pipeline, not a general support queue.
fluorocell@yashara.usWe provide structured 30-day evaluation access for qualified research teams. To make the evaluation useful rather than generic, we configure it to your specific cell type and imaging modality where possible. Tell us your context in the message field.
Product demonstrations are technical conversations, not sales presentations. If you tell us your cell line and imaging setup in advance, we come prepared with relevant segmentation examples — not a deck designed for a general audience.
We respond to all substantive inquiries within two business days. Replies come from a scientist, not a CRM sequence.