segmentation-quality-control 0.1.1

MIDL 2024 paper: Efficiently correcting patch-based segmentation errors to control image-level performance in retinal images

Efficiently correcting patch-based segmentation errors to control image-level performance

Experiments for MIDL 2024 paper.

Installation

^{Requires Python >= 3.9}

Development setup with uv (recommended)

Clone the repository:

git clone https://github.com/berenslab/MIDL24-segmentation_quality_control.git
cd MIDL24-segmentation_quality_control
uv sync --extra dev
source .venv/bin/activate  # Windows: .venv\Scripts\activate

Development setup without uv

Alternatively, use Python's built-in venv:

git clone https://github.com/berenslab/MIDL24-segmentation_quality_control.git
cd MIDL24-segmentation_quality_control
python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -e ".[dev]"

From PyPI (recommended for use as a dependency)

Install the released package:

pip install segmentation_quality_control

PyPI wheels include the Python code only. Ensemble checkpoints (~360 MB) are not bundled; see Model weights below.

From GitHub (development version)

Install the latest development version without cloning:

pip install git+https://github.com/berenslab/MIDL24-segmentation_quality_control

Model weights

The FR-UNet ensemble uses five checkpoints: FRUNet_0.pth, ..., FRUNet_4.pth.

Where weights are resolved

When loading models, the package resolves the checkpoint directory in this order:

1. Custom path: pass models_dir=... to ensure_models_dir().
2. Local development checkout: if all five files are present in the repository's trained/ folder, that directory is used. This applies to editable installs (pip install -e .) and cloned repositories.
3. User cache (OS default for PyPI and other installs):

OS	Default cache directory
Linux	~/.cache/segmentation_quality_control/trained/
macOS	~/Library/Caches/segmentation_quality_control/trained/
Windows	%LOCALAPPDATA%\segmentation_quality_control\trained\

Automatic download

If weights are missing from the resolved location (such as when using as a pypi dependency), they are downloaded over HTTPS from GitHub (trained/) into the cache directory:

from segmentation_quality_control.utils import ensure_models_dir

models_dir = ensure_models_dir()  # downloads on first call if needed

If Python's https-download mechanism fails, the package falls back to using git when it is available on PATH.

Manual download

If the automatic download fails, fetch the files directly:

mkdir -p ~/.cache/segmentation_quality_control/trained
wget -O ~/.cache/segmentation_quality_control/trained/FRUNet_0.pth \
  https://raw.githubusercontent.com/berenslab/MIDL24-segmentation_quality_control/main/trained/FRUNet_0.pth
# repeat for FRUNet_1.pth … FRUNet_4.pth, or use curl -L -o ...

Alternatively, use a sparse git checkout:

git clone --depth 1 --filter=blob:none --sparse \
  https://github.com/berenslab/MIDL24-segmentation_quality_control.git \
  ~/tmp/segmentation_weights
cd ~/tmp/segmentation_weights
git sparse-checkout set trained
git checkout main
# then copy trained/FRUNet_*.pth into your cache or repo trained/ directory

Checkpoint format

• Weights are plain PyTorch checkpoints (not bundled in PyPI wheels).
• Keys: arch, epoch, state_dict, optimizer, config
• Loading no longer requires the legacy bunch package.
• To save new checkpoints after retraining, use segmentation_quality_control.utils.checkpoint.save_checkpoint, to store them in the bunch-less format.

Reproduce the paper results

Instructions to reproduce the experiments (development setup):

0. Install the package as above (uv sync --extra dev or pip install -e ".[dev]").
1. Download the FIVES dataset (Jin et al., 2022) and register its path in paths.yaml.
2. Execute 0_pass_forward.ipynb to pass the data through the ensemble and log the outputs. In a cloned repository, weights are read from trained/.
3. Calibrate with temperature scaling on the validation set with 1_calibrate.ipynb.
4. Compute $\widehat{DSC}$ and $\widehat{DSC}_{\text{corr}}$ in 2_estimate_DSC.ipynb.
5. Reproduce figures with the respective scripts.

References

Jin, Kai, et al. "Fives: A fundus image dataset for artificial Intelligence based vessel segmentation." Scientific Data 9.1 (2022): 475.

Paper and Citation

The paper (MIDL 2024 oral):

• Köhler et al., Efficiently correcting patch-based segmentation errors to control image-level performance in retinal images (2024)
• PDF at OpenReview

Citation:

@inproceedings{
koehler2024efficiently,
title={Efficiently correcting patch-based segmentation errors to control image-level performance in retinal images},
author={Patrick K{\"o}hler and Jeremiah Fadugba and Philipp Berens and Lisa M. Koch},
booktitle={Medical Imaging with Deep Learning},
year={2024},
url={https://openreview.net/forum?id=DDHRGHfwji}
}

Note

This package has been bundled into the Fundus Image Toolbox for convenient application next to other useful retinal fundus imaging tools.