🧬 Computer Vision for Bacteria Counting

🚀 Introduction

This project is an extension of the 2022 DeepBacs paper, focusing on enhancing cell recall in dense, high-contrast regions. Our approach combines advanced deep learning methodologies to address the challenges posed in densely populated, high-contrast images.

Key Highlights:

• Data: Available in the Data folder on SCC, labeled according to figures in the original DeepBacs paper.
• Approach:
  • Pre-task training with a Convolutional Neural Network (CNN) to extract feature representations.
  • A Transformer-based segmentation model that utilizes the pre-task CNN features and the original grayscale image to produce precise segmentations.
• Goal: Improve upon StarDist2D's generalist model and the methodologies in DeepBacs for segmenting cells in challenging environments.

---

🛠️ Setup Instructions

1. Dependencies

Ensure the following libraries are installed:

!pip install numpy matplotlib random scikit-image torch pillow einops opencv-python

Alternatively, list them in a requirements file:

pip install -r requirements.txt

2. Primary Results

To reproduce our results, follow these steps:

• Run the Transformer-CNN pipeline:

  jupyter notebook TransformerCNN_masks.ipynb

  • Input: Raw grayscale images of S. Aureus dataset (available at Zenodo).
  • Output:
    • Individual instance masks for each bacterium.
    • Combined binary masks (sum across all instance masks).

• Augmented Dataset Analysis:

  jupyter notebook Augmented_training_pipeline.ipynb

3. Output Example

Below is an example of the segmentation output, showcasing individual bacterial masks:

---

📊 Work Log

Date	Task	Contributor(s)
10/10/2024	Uploaded StarDist demo for initial analysis.	Isaac
11/19/2024	Added Transformer framework class.	Omar
11/20/2024	Integrated StarDist2D generalist model and updated the Data folder.	Isaac
11/20/2024	Added CNN framework class.	Mohi
12/06/2024	Integrated pre-trained ResNet for CNN features.	Mohi / Berk
12/07/2024	Initial version of the Transformer-based model.	Isaac / Omar
12/07 → 12/15	Developed multiple Transformer model variations.	Everyone
12/15/2024	Finalized model selection, cleaned GitHub structure, and organized main notebooks into folders.	Everyone

---

🧪 Results Summary

• Target Metric: Improved recall compared to StarDist2D.
• Approach: Transformer + CNN pre-task framework.
• Key Focus: Dense, high-contrast regions with challenging segmentation cases.

---

🌟 For any questions, please feel free to contact our team or refer to the detailed notebooks in the Notebooks folder.

---