Overview  ·  Installation  ·  Usage  ·  Outputs  ·  Agreement study  ·  Documentation  ·  Citation

CelloType is a local web application for the quantitative analysis of multichannel fluorescence microscopy images. It takes a dataset from raw files to per-object measurements in a single guided workflow, and runs entirely on the researcher's machine.

Overview

_{Workflow. Orange outlines mark decisions taken by the researcher; grey outlines, automated steps. Masks are reviewed and edited before anything is measured (dashed red: back to segmentation).}

Mechanical steps are automated; experimental decisions stay with the researcher and are recorded next to the results.

• Input. CZI, TIFF (OME or ImageJ) and PNG/JPEG at native bit depth, with channel names and pixel size read from the metadata.
• Segmentation. Cellpose-SAM for nuclei, cells or other objects on any channel, on CUDA or Apple Silicon GPUs with a CPU fallback; optional StarDist for nuclei.
• Review. Every mask can be inspected and edited in the browser before measurement.
• Quantification. 18 ImageJ-style measurements per object on every channel with the same masks, sizes in µm, optional background subtraction, and border objects excluded by default.
• Analysis. Distributions, co-expression and positivity against a control condition.
• Reproducibility. Each table can be exported with a JSON record of its parameters and the final label masks.

Architecture

_{At each stage the interface sends a request to a FastAPI endpoint (under /api), a Python module does the work, and its output feeds the next stage. Details in docs/ARCHITECTURE.md.}

Installation

Requires Python 3.10 to 3.12. A GPU is optional but strongly recommended (NVIDIA CUDA or Apple Silicon MPS). The Cellpose-SAM weights (about 1.2 GB) are downloaded to ~/.cellpose on first use.

git clone https://github.com/abrilcode/CelloType.git
cd CelloType
python3 -m venv .venv
source .venv/bin/activate          # Windows (PowerShell): .venv\Scripts\Activate.ps1
pip install --upgrade pip
pip install -r requirements.txt

On an NVIDIA GPU, install the CUDA build of PyTorch that matches the driver before requirements.txt. StarDist is optional: pip install stardist tensorflow.

Usage

uvicorn app:app --host 127.0.0.1 --port 8010

Open http://127.0.0.1:8010 and choose Single image for one field, or Multiple images for a dataset organised into experimental conditions. The interface leads through Raw data, Channels, Segment, Measure and Analysis.

_{Segment step: 178 nuclei on the DAPI channel with Cellpose-SAM; the 20 that touch the image border are excluded and 158 are measured.}

Outputs

File	Contents
measurements.csv	One row per object and channel, with the selected measurements.
<table>_parameters.json	Inputs, settings, mask edits and software versions, and whether the result can be reproduced from the parameters alone.
Masks	The final label image as TIFF, or one TIFF per field in a ZIP.
cells_positive.csv	Positivity per object and marker (multiple images).
Plots	PNG figures of each analysis view.

Columns and the parameter record are described in docs/OUTPUTS.md.

Agreement study

Nuclear segmentation and per-nucleus readouts were compared with a published StarDist workflow, Multi-Channel Nuclear Analysis (Waisman, 2025), on five datasets: 78 fields and 21,385 reference nuclei, both mask sets measured with the same code.

_{(a) Dice, F1 at IoU ≥ 0.5 and matched IoU per dataset, against the 0.85 threshold. (b) Nuclei detected by CelloType against the reference; the band is ±5 %. (c) Pearson r of per-nucleus mean intensity for each marker, on matched nuclei.}

Every dataset met the thresholds set before the comparison (Dice 0.948 to 0.964, F1 0.942 to 0.965, count r ≥ 0.991, intensity r ≥ 0.997). This is agreement between two automatic methods, not accuracy against manual annotation; on these images Cellpose-SAM was 15 to 37 times slower than StarDist. Method, per-dataset results and timing are in docs/VALIDATION.md.

Documentation

Document	Contents
docs/VALIDATION.md	Agreement study: method, thresholds, per-dataset results and timing
docs/OUTPUTS.md	Output files, columns, measurements and the parameter record
docs/ARCHITECTURE.md	Modules by layer and design notes
CHANGELOG.md	Changes since the version evaluated in the thesis (v1.0.0)

Data and privacy

Images are processed locally and never uploaded. Microscopy images and patient-derived data must not be committed to this repository; the .gitignore excludes common microscopy formats. CelloType is research software and is not intended for diagnostic use.

Citation

CelloType was developed by Abril Herrlein as the final project of the Bioengineering degree at the Instituto Tecnológico de Buenos Aires (ITBA), supervised by Dra. María Andrea Camilletti. Citation metadata is also available through "Cite this repository" (CITATION.cff).

@thesis{herrlein2026celloType,
  author      = {Herrlein, Abril},
  title       = {Desarrollo de una plataforma web para segmentaci{\'o}n y cuantificaci{\'o}n
                 de im{\'a}genes de fluorescencia multicanal},
  type        = {Proyecto final de carrera, Bioingenier{\'i}a},
  institution = {Instituto Tecnol{\'o}gico de Buenos Aires},
  year        = {2026}
}

Results produced with CelloType should also cite the segmentation model used: Cellpose-SAM (Pachitariu, Rariden and Stringer, bioRxiv 2025, doi:10.1101/2025.04.28.651001) and Cellpose (Stringer et al., Nature Methods 2021, doi:10.1038/s41592-020-01018-x), or StarDist (Schmidt et al., MICCAI 2018, doi:10.1007/978-3-030-00934-2_30). The reference workflow of the agreement study is Waisman, microPublication Biology 2025, doi:10.17912/micropub.biology.001590.

License

Released under the MIT License.