MAPS: Machine-Assisted Phenotype Scoring

Most cell biology experiments involve scoring cell-level phenotypes from microscopy images. This process is very labour-intensive as it is traditionally done manually.

MAPS is designed to automate the processing and interpretations of large amounts of microscopy images, but is adaptable to other types of images.

MAPS is built built around using Microsoft's Azure CustomVision.ai which has an intuitive graphical interface for model traning.

Detailed implementation guide is published to protocols.io.

Steps:

Beofre we start: Input images should be converted to TIFF files of microscopy images containing Red, Green and Blue (DAPI stained nuclei) channels.

1. Pre-processing: This is a quality control step to remove blurry (out-of-focus images). Also some basic histogram corrrection is applied.

2.1 Cell detection: An object detection model built on Azure Custom Vision is used to detect individual cells from images. Bounding boxes for detected cells are visualized. Each cell is cropped from the images.

2.2. Training augmentation: Augment your training data with 14 different image transformations. The original bounding box coordinates are preserved. Images are re-uploaded to Azure for re-training. Prereq: bbox_util.py

3. Data visualization: Generate cell galleries to help you visualize phenotypes. First, image features are extracted by parallel stacking of 3 convolutional layers, then t-SNE (with PCA initiallization, Kobak & Berens Nat. Comm. 2019) and spectral clustering are applied.

4. Phenotype scoring: A classification model built on Azure is used to classify the cells into different bins, based on the general classes of localizations discovered in step 2.2.

Note: Because different experiments involve different cell types and proteins, users should build their own object detection and classification models on Azure.

Visit this quick start guide on how to build your first object detection model (no coding required! yay!).

Read this how-to to learn about using your Azure Custom Vision model with the Python SDK.