Find nuclei
Predict foreground structure and separate it into individual integer-labelled instances.
BioNucleiBoundary U-NetBBBC039v1 baseline
Know what the model is doing before you trust the result.
BioNuclei is a research system for nuclear instance segmentation in fluorescence bioimaging. Here you can see the documented training setup, input assumptions, measurements, limitations and the proposed adaptive-agent layer.
01 · Task
What does one uploaded image become?
The current core pipeline answers a focused question: where are the individual nuclei? It then derives measurable properties of those predicted instances.
Measure morphology
Compute geometric properties such as area, centroid and bounding box; extended reports add shape descriptors.
Measure intensity
Compute per-nucleus fluorescence statistics on the selected image plane when enabled.
02 · Model
The current AI is a Boundary U-Net.
It is a convolutional neural network. The scientific Python implementation performs the actual image inference and metrics; an agent layer does not replace it.
1-channel
fluorescence
fluorescence
→
Boundary U-Net
CNN
CNN
→
background
interior
boundary
interior
boundary
→
instance mask
measurements
measurements
Architecture
Boundary-aware U-Net with 1 input channel and 3 output classes in the documented baseline.
Post-processing
Predicted foreground is converted to separate instances with deterministic connected-component labeling.
No online retraining
An uploaded user image is an inference input. The current prediction path does not update the checkpoint.
03 · Training
What was the documented model trained on?
The values below are the repository's BBBC039v1 baseline configuration.
| Parameter | Value |
|---|---|
| Dataset | BBBC039v1 |
| Input | 1 channel, uint16, 520 × 696 |
| Classes | Background / nuclear interior / boundary |
| Base channels | 32 |
| Seed | 42 |
| Epochs | 100 |
| Batch size | 8 |
| Optimizer | AdamW |
| Learning rate | 3 × 10⁻⁴ |
| Weight decay | 1 × 10⁻⁵ |
| Loss | Cross-entropy + boundary weight 2.0 + Dice weight 1.0 |
| Augmentation | Horizontal flip, vertical flip, 90° rotation |
Important: these settings describe the documented baseline recipe. A released checkpoint must carry its own configuration/provenance. Do not assume every checkpoint is identical to this baseline.
04 · Input
What should a user upload?
TIFF
A 2-D fluorescence TIFF can be passed directly to the current 2-D inference implementation.
Nikon ND2
The community service can select channel, time, Z and field/position, then extract a defined 2-D plane for the current model.
Quality matters
Supported format is not the same thing as validated biology. Staining, optics, magnification, saturation, noise and domain shift can affect reliability.
05 · Output
What will the scientist receive?
Segmentation
Integer-labelled mask and visual overlay showing predicted nuclei.
Nucleus parameters
Instance ID, area, centroid and bounding-box coordinates; extended analysis adds morphology descriptors.
Intensity
Per-nucleus fluorescence statistics when intensity analysis is enabled.
Reports
Nuclei, morphology and intensity JSON/CSV summaries produced by the analysis service.
Provenance
Input identity, selected ND2 plane, model/checkpoint and execution metadata.
Download
A complete ZIP bundle for downstream analysis and archiving.
06 · Adaptive AI
Yes: adaptive agents can make BioNuclei much stronger.
The safest design is a layered system. Agents can reason about the request and route the image to validated CNNs, while deterministic code remains responsible for measurements. That gives us flexibility without turning the scientific result into an opaque LLM answer.
Acquisition agent
Reads ND2/TIFF metadata, identifies channels/planes and checks whether the input is compatible with an enabled workflow.
Model router
Selects among released, validated CNNs. If no model fits the input domain, it refuses to guess.
Quality agent
Runs explicit quality diagnostics and flags saturation, low signal, unusual object statistics or other implemented failure indicators.
Analysis planner
Maps a user's biological question to available validated modules: nuclei, morphology, intensity and later specialist tasks.
Report agent
Assembles visual results, tables, warnings and provenance into a readable report while preserving raw machine-readable evidence.
Learning loop
Opt-in research contributions can be curated, annotated and used in a separate training pipeline. They should never silently modify the live model.
Future multi-model architecture: one acquisition may route to a nuclei CNN, a morphology workflow, a fluorescence/intensity model, and later specialized models for cells, membranes, organelles or tracking—each with its own scientific validation gate.