Glossary

Definitions of key terms used in the PhenoMe documentation, in plain language.

A–D

Centroid

The average (mean) position of a group of points in embedding space. When computing “distance to centroid,” we measure how far each image’s embedding is from the mean embedding of a reference group (e.g., controls).

Channel

A single color or wavelength layer in a microscopy image. A 3-channel fluorescence image might have DAPI (nucleus), GFP (protein), and mCherry (membrane). The pipeline can process each channel separately (split mode) or combine them as RGB (combined mode).

Channel mode

How multi-channel images are handled. Split mode: each channel is processed independently; embeddings are concatenated. Combined mode: channels are stacked as RGB and processed together. Use split for fluorescence; use combined for brightfield or natural color.

Checkpoint

A saved state of your analysis (embeddings, metadata, properties) stored in HDF5 format. Enables resuming after interruption and sharing results.

Classical properties

Hand-crafted numerical features (area, intensity, eccentricity, etc.) computed from images and masks. Interpretable and used as the “vocabulary” to explain what embeddings encode.

CLS token

In Vision Transformers (ViTs), a special token whose output represents a summary of the entire image. Used as the embedding vector.

Correlation coefficient (r)

A number between -1 and 1 measuring how two variables move together. Positive r: both increase together. Negative r: one increases as the other decreases. Near 0: little linear relationship. Used to link embedding dimensions to classical properties for interpretability.

Dimensionality reduction (visualizing patterns)

Reducing complex high-dimensional information (like 768 numbers describing an image) into simpler representations (2 or 3 numbers) so we can plot them on a map. Methods: PCA, t-SNE, UMAP.

DINOv2

A self-supervised vision model by Meta AI. Learns visual features without labels. Default example model in the pipeline.

Distance metric

How “far apart” two embeddings are. Euclidean: straight-line distance. Cosine: angle between vectors (scale-invariant).

E–M

Embedding

A fixed-size numerical vector that represents an image. Produced by a deep learning model; captures visual characteristics in a compact form.

Metadata

Information about each image: file path, condition, drug, timepoint, replicate, etc. Extracted from file paths or CSV.

Model-agnostic

The pipeline works with any vision model that produces embeddings (DINOv2, CLIP, BioCLIP, custom models). No lock-in to a single architecture.

Morphometry

Quantitative measurement of shape and form. Classical morphometry uses hand-crafted features (area, perimeter, eccentricity); the pipeline also uses deep embeddings.

P–R

PCA (Principal Component Analysis)

Linear dimensionality reduction. Finds directions of maximal variance. Fast, deterministic. Good for a quick overview.

Phenotypic distance

A scalar value indicating how different an image is from a reference group (e.g., controls). Higher distance = more phenotypically different.

Phenotyping

Characterizing observable traits (phenotypes) of biological samples, e.g., cell morphology, fluorescence patterns, response to treatments.

Property preset

A predefined set of classical properties: none, intensity, shape, basic, standard, or complete. See Select properties — Presets and Property interpretation.

Reference group

The set of images used as baseline (e.g., untreated controls). Defined by reference_filters when computing phenotypic distances.

Report

A standalone HTML file with interactive plots, distance analysis, clustering, correlations, and image galleries. Shareable without re-running the pipeline.

S–Z

Split mode

See Channel mode. Each channel processed independently.

t-SNE

Non-linear dimensionality reduction. Preserves local structure; good for clustering visualization. Can be stochastic (results vary with seed).

UMAP

Non-linear dimensionality reduction. Balances global and local structure; often faster than t-SNE. Default for many exploratory workflows.

Zero-shot

Using a pre-trained model without training or fine-tuning on your data. DINOv2 and similar models work out-of-the-box for phenotyping.