PolusAI/theia

A scalable deep learning based bleed through correction algorithm.

Theia (v0.1.2)

Bleed-Through Correction in Fluorescent Microscopy Images

Installation

Linux

• From PyPI: pip install "theia-py==0.1.2"
• Using poetry: poetry add theia-py@0.1.2

MacOS (Apple Silicon)

1. Install xcode command line tools:
   • xcode-select --install
2. Install miniforge:
   • Instructions
3. Disable the base conda environment from activating by default:

• conda config --set auto_activate_base false

4. Install poetry:

• curl -sSL https://install.python-poetry.org | python3 -
• Add poetry to your path using the hint provided by the poetry installer.

5. Create a conda environment:

• conda create -n theia python=3.9
• conda activate theia

6. Install tensorflow dependencies:

• conda install -c apple tensorflow-deps

7. Install using poetry:

• poetry add theia-py@0.1.2

Windows

1. Re-evaluate your life choices.
2. Install Linux.
3. Read this document from the beginning.

Usage

Here is a simple example of how to use Theia to correct a set of images. For a more detailed example, see the streamlit app in examples/rxrx.py.

import theia

# Load training and validation images
train_images = ...
valid_images = ...

# Make a tile-generator for feeding the model
train_generator = theia.TileGenerator(train_images, tile_size=(256, 256), normalize=False)
valid_generator = theia.TileGenerator(valid_images, tile_size=(256, 256), normalize=False)

# Build the model
model = theia.models.Neural(
    num_channels=...,
    channel_overlap=1,
    kernel_size=5,
    alpha=1,
    beta=1,
    tile_size=256,
)
model.early_stopping(
    min_delta=1e-3,
    patience=4,
    verbose=1,
    restore_best_weights=True,
)
model.compile(optimizer="adam")
model.fit_theia(
    train_gen=train_gen,
    valid_gen=valid_gen,
    epochs=128,
    verbose=1,
)

# Get the transformer
transformer = model.transformer

# Correct and save the images
for image in train_images + valid_images:
    corrected_image = transformer.transform(image, remove_interactions=True)
    with open("<path-to-image>", "w") as f:
        ...