Warning
Work in progress, not intended for public use. This repository is the research record behind the MegaSeg paper: the experiments, the benchmark and the code that makes the figures. It is shared for transparency and is not maintained as software. Paths point to Allen Institute storage, the model checkpoints and most of the data are not included, and the code is not expected to run outside our environment. We do not provide support, and issues and pull requests may not get a response.
To segment your own images or fine-tune the model, use SubCellSeg. It is the released, maintained version of the MegaSeg model and installs with one command. MegaSeg was the name of the model during development.
Multi-structure segmentation for fluorescently-tagged cellular structures. This repository contains classical segmentation workflows, ML-based segmentation experiments, and a comprehensive evaluation benchmark comparing MegaSeg (multi-structure) models against individual (specialized) models.
MegaSeg/
├── segmenter-classic/ # Classical segmentation notebooks (18 notebooks)
├── segmenter-ml/ # ML training configs and experiment tracking
├── benchmark/ # Evaluation benchmark (14 structures, 4 backbones)
├── data/
│ ├── variance-100x/ # Per-structure CSVs (train/val/test splits)
│ └── standardized_test/ # Held-out test_new dataset
└── Results_Figure_Tracking/ # Figure tracking CSV
Classical (non-ML) segmentation pipelines implemented as Jupyter notebooks. Covers structure-specific segmentation using traditional image processing methods.
ML-based segmentation training configurations for individual and multi-structure models. Contains Hydra configs for training with cyto-dl.
Systematic evaluation comparing MegaSeg multi-structure models (DynUNet, AttentionUNet, SwinUNETR, UNet) against 14 individual specialized models across multiple evaluation paradigms. Includes config generation, batch evaluation runners, quantitative metrics computation, and visualization.
See benchmark/README.md for full details.
Test set CSVs for 14 structures live in data/variance-100x/{STRUCTURE}/ with train.csv, val.csv, and test.csv splits. The held-out evaluation dataset is at data/standardized_test/test_new.csv.
Structures: ACTN1, ATP2A2, CTNNB1, DSP, GJA1, MYH10, NPM1, NUP153, RAB5A, SEC61B, SON, ST6GAL1, TJP1, TOMM20
# Clone with submodules (benchmark uses cyto-dl as a submodule)
git clone --recurse-submodules git@github.com:AllenCell/MegaSeg.git
# Or initialize submodules after cloning
git submodule update --init --recursive
# Install benchmark dependencies
cd MegaSeg/benchmark
uv sync