Overview

The Segmenter ML plugin is based on the iterative deep learning workflow of the Allen Cell & Structure Segmenter and CytoDL, both are developed at the Allen Institute for Cell Science:

  • Allen Cell & Structure Segmenter is a Python-based open-source toolkit for 3D segmentation of intracellular structures in fluorescence microscope images

  • CytoDL is a codebase unifying deep learning approaches for understanding 2D and 3D biological data as images, point clouds, and tabular data

The Allen Cell Segmenter ML plugin has 3 main modules: Curation, Training, and Prediction.

1. Curation

  • This module assists user in curating training dataset through manual sorting, excluding, & merging image data

  • Data curation step is important as a model’s performance is directly tied to the training data’s quality

a. Sorting

  • Review raw images and corresponding segmentations

  • Select only the high-quality images to be used as training data

b. Excluding

  • Select regions within an image to exclude from training

  • Maximize usable data by not rejecting the entire image but just the problem regions

c. Merging (overwriting)

If a raw image has two segmentations of the same cellular structure produced by different algorithms to target specific morphologies, Merging allows users to:

  • Select one of the segmentations to be the base segmentation

  • Draw shape(s) to indicate region(s) from the supplementary segmentation to overwrite the same region(s) in the base segmentation, effectively merging the two segmentations into a single ground-truth segmentation to be used for training

  • For 3D images, these regions are applied through all z-slices


2. Training

images/training.png

This module allows users to train an ML 2D or 3D segmentation model from scratch or fine-tune (iteratively) an existing 2D or 3D segmentation model**–whether their own or a pre-trained model provided by us–using their own data.


3. Prediction

images/prediction.png

This module allows users to apply the trained ML model from the previous step, or a pre-trained model, to generate segmentation predictions on raw images that the model has not previously seen.