Astrocyte-detection

Astrocytes are star-shaped glial cells in the nervous system that play a crucial role in maintaining brain homeostasis and are implicated in various neurological disorders. This machine learning pipeline enables researchers and clinicians to automatically detect astrocytes from microscopy images of human brain tissue, improving the speed and accuracy of analysis in neurobiological studies, potentially supporting neuropathological disease diagnosis, progression tracking, and large-scale post-mortem analysis in research and clinical settings.

Overview

This repository provides a full pipeline for astrocyte segmentation from stained brain scan tiles using a self-supervised SimCLR-based encoder and a modified U-Net with a ResNet-18 backbone. SimCLR was chosen to address the scarcity of annotated data at the time, enabling the model to learn meaningful features from over 100,000 unlabeled 512×512 tiles. The encoder was then fine-tuned on only 200 semi-automatically segmented tiles generated via a rule-based pipeline included in this repo.

Features

• Self-supervised SimCLR encoder trained on large unlabeled dataset of brain tiles

• Modified U-Net using the pretrained encoder with a ResNet-18 backbone

• Semi-automated annotation tool for generating segmentation masks

• Pretrained weights and configs included for reproducibility

• Easy retraining support with custom data for both encoder and U-Net

Run Locally

Clone the project

  git clone https://github.com/raisaaajose/astrocyte-detection.git

Go to the project directory

  cd astrocyte-detection

Install dependencies

  pip install -r requirements.txt

Run pretrained U-Net on test images

  python unet/predict.py --input_dir path/to/images --weights pretrained/unet.pth

Datasets

Unlabeled: 100,000+ 512×512 brain tile images used to train the encoder were generously provided by the Sudha Gopalakrishnan Brain Centre (SGBC), IIT Madras, as part of their ongoing research in astrocyte morphology.

Labeled: 200 tiles with segmentation masks created semi-automatically using the rule-based labeling pipeline

Labeled Masks

To avoid the costs of manual annotations, a simple rule-based pipeline was used to generate rough segmentation masks that can be tweaked manually if needed. A series of median blurring and erosion operations were applied on the binary mask to identify potential cell centroids. These centroids were then used as seed points to apply flood-filling on the same mask, producing complete semantic segmentations. This approach gave us around 200 semi-automated masks, which were good enough to fine-tune the model effectively.

Suggestions for Users

• Start by using the pretrained models to evaluate performance on your dataset.

• For domain adaptation, retrain the SimCLR encoder using your own unlabeled brain tiles.

• Use the semi-automated rule-based tool to quickly generate initial masks for small datasets.

TODO / Future Work

• Add support for larger tile sizes

• Add support for Docker

• Improve mask refinement in the rule-based pipeline

• Integrate interactive annotation support

• Integrate CLI flags for directory paths

🤝 Contributions

Contributions are welcome! Please see the CONTRIBUTING.md for guidelines.

Acknowledgements

• A Simple Framework for Contrastive Learning of Visual Representations (SimCLR)
• SimCLR Implementation in PyTorch by Spijkervet
• Sudha Gopalakrishnan Brain Centre (SGBC), IIT Madras
• Classification of Microglial Morphological Phenotypes Using Machine Learning – reference for pipeline design inspiration