Pico-Algae Detection using Deep Learning

Overview

This project detects and counts pico-algae cells in microscopy images using a Faster R-CNN detector. It is useful for fast, repeatable analysis of samples where manual counting is slow and error-prone. The pipeline works on paired microscope images (*_og.png and *_red.png) and predicts cell-level bounding boxes with class labels. The repository includes training, tuning, and inference utilities for reproducible experiments.

Example Results

Input image:

Detections with bounding boxes:

Pipeline Overview

Image -> preprocessing -> object detection -> counting

1. Load microscopy image pair (og + red).
2. Preprocess and resize to model target resolution.
3. Run Faster R-CNN for small-object detection.
4. Filter predictions by confidence and count detections.

Model

The detector is a 6-channel Faster R-CNN (ResNet50-FPN) that fuses og and red images. Training uses annotated pico-algae microscopy data with bounding boxes and four foreground classes (EUK, FE, FC, colony). The setup is tuned for dense small-object detection where many tiny cells appear in one frame.

Installation

pip install -r requirements.txt
pip install -e .

Usage

Single-image inference (package entrypoint):

python -m pico_algae.inference --image examples/input_images/test.jpg

Single-image inference (console script after editable install):

pico-algae-infer --image examples/input_images/test.jpg

Batch/folder inference (existing pipeline):

python scripts/predict_frcnn.py --images_dir data/raw/images_og --ckpt runs/train_run01/checkpoints/best_mae.pt --out_dir runs/predict_run --predict_yaml src/configs/predict_frcnn.yaml

Training:

python scripts/train_frcnn.py --index_csv data/processed/dataset_2048x1500_webp/index.csv --out_dir runs/train_run --train_yaml src/configs/train_frcnn.yaml

Root CLI wrapper:

python main.py infer --image examples/input_images/test.jpg

Results

• Classes detected: EUK, FE, FC, colony.
• Best post-processing tuning result (runs/tuning/post_best_mae/tuning_post_results.csv): mean_count_mae=2.4239, std_count_mae=0.5050 across 5 folds.
• Dense-scene challenge: small, overlapping cells can increase count error in crowded regions, so threshold and NMS tuning materially affect counting quality.

Project Structure

pico-algae-detection/
├── README.md
├── requirements.txt
├── LICENSE
├── configs/
│   └── model_config.yaml
├── docs/
│   └── pipeline_diagram.png
├── examples/
│   ├── example_input.png
│   ├── example_detection.png
│   ├── input_images/
│   └── detection_results/
├── notebooks/
│   └── exploration.ipynb
├── scripts/
│   ├── prep/
│   ├── train_frcnn.py
│   ├── predict_frcnn.py
│   └── inference.py
├── src/
│   ├── data/
│   ├── models/
│   ├── train/
│   ├── inference/
│   ├── utils/
│   └── configs/
└── models/

Repo Conventions

• src/: importable implementation code (model, training logic, inference utilities, helpers).
• scripts/: runnable entrypoints and operational scripts.
• scripts/prep/: dataset preparation and one-off data processing utilities.

Technologies

• Python
• PyTorch
• OpenCV
• NumPy

Folder Guide

• src/train: training loop, optimization, scheduler, and evaluation code.
• src/inference: prediction and visualization utilities.
• src/models: Faster R-CNN model builders and checkpoint utilities.
• src/configs: YAML configs for train, predict, and tuning.
• src/utils: shared helper functions for I/O, seeding, logging, and box ops.
• scripts/prep: dataset preprocessing and annotation conversion scripts.