ReScale4DL: Balancing Pixel and Contextual Information for Enhanced Bioimage Segmentation

A systematic approach for determining optimal image resolution in deep learning-based microscopy segmentation, balancing accuracy with acquisition/storage costs. Following this approach, researchers can improve the sustainability and cost-effectiveness of bioimaging studies by reducing data and computing needs while optimising microscopy techniques.

Key Features

• Resolution simulation: Rescale images and their respective annotations (upsample and downsample)
• Segmentation evaluation: Compare performance across resolutions using:
  • Mean Intersection-over-Union (IoU)
  • Morphological features
  • Potential throughput
  • Personalised metrics
• Visualization tools: Generate comparative plots and sample outputs

Installation

ReScale4DL is available as a Python package through pip. Activate your conda environment or create one:

conda create -n rescale4dl "python<=3.12"
conda activate rescale4dl

Install ReScale4DL with pip:

pip install rescale4dl

Manual installation

Manual installation using the GitHub repository

git clone https://github.com/HenriquesLab/ReScale4DL.git
cd rescale4dl
conda create -n rescale4dl "python<=3.12"
conda activate rescale4dl
python -m pip install .

Usage

1. Image Rescaling

Notebook: Rescale_Images.ipynb

2. Segmentation Analysis

Notebook for 2D images: Evaluate_Segmentation_2D.ipynb

Notebook for 3D images: Evaluate_Segmentation_3D.ipynb

Notebook to reproduce ReScale4DL results on real acquisitions: Evaluate_Segmentation_real_acquisirions.ipynb

3. Rescale and crop to test foundation models

Notebook: Rescale_Foundation_Models.ipynb

How to use demo

https://github.com/user-attachments/assets/ff67aee1-9e7b-45c1-93f5-7e4c66c6c860

https://github.com/user-attachments/assets/74defb21-9470-4cf3-b993-d6710cfb22c4

Additional DL resources for microscopy:

The deep learning networks presented in the ReScale4DL paper were trained using the following platforms:

• ZeroCostDL4Mic: A Google Colab-based no-cost toolbox to explore Deep Learning in Microscopy
• DL4MicEverywhere: Docker-based implementation bringing the ZeroCostDL4Mic experience for local deployment

For detailed hyperparameter settings and training configurations, please refer to Table 1 in our bioRxiv preprint.

Scripts

ReScale4DL provides a set of functionalities to quickly analyse your images and find an optimal pixel size:

1. Rescaling the images in the path by donwsampling with a factor of 2 and 3, and by upsampling with a factor of 2:

rescale4dl.batch.process_all_datasets(“/path/data”, [2,3], [2], [1], modes=[“mean”])

2. Analyse the segmentation results for different scaling factors in 2D:

rescale4dl.analyse(“/path/data”) 

3. Analyse the segmentation results for different scaling factors in 3D:

rescale4dl.analyse(“/path/data”,
        is_3d=True,
        run_per_object_stats = False, # True for Instance Segmentation, False for Semantic or Binary Segmentation
        save_images = False, # True to save images of the segmentation examples and data distributions, False to skip saving images and saving some memory
        sampling_dir_list = None)

Contributing

We welcome contributions through:

• Issue reporting
• Pull requests

License

MIT License - See LICENSE for details

How to cite this work

Ferreira, M.G., Saraiva, B.M., Brito, A.D., Pinho, M.G., Henriques, R. and Gómez-de-Mariscal, E., ReScale4DL: Balancing Pixel and Contextual Information for Enhanced Bioimage Segmentation. bioRxiv, pp.2025-04, (2025) https://doi.org/10.1101/2025.04.09.647871

@article{ferreira2025rescale4dl,
  title={ReScale4DL: Balancing Pixel and Contextual Information for Enhanced Bioimage Segmentation},
  author={Ferreira, Mariana G and Saraiva, Bruno M and Brito, Ant{\'o}nio D and Pinho, Mariana G and Henriques, Ricardo and G{\'o}mez-de-Mariscal, Estibaliz},
  journal={bioRxiv},
  pages={2025--04},
  year={2025},
  publisher={Cold Spring Harbor Laboratory},
  URL = https://doi.org/10.1101/2025.04.09.647871
}