soma¶
soma is a modular framework to streamline computational pathology research.
Data and evaluation are fixed scaffolding; the encoder and decoder are the two swappable axes — change one in a single line and keep the rest.¶
It provides a unified API to go from a dataset of slides and labels to a full, reproducible result report. Along the way, it makes it easy to sweep core design choices such as preprocessing (spacing, field-of-view), encoding (foundation models), and aggregation (MIL) so you can quickly find the strongest configuration for your data.
Under the hood, it builds on two open-source projects of mine:
soma supports tile, slide, and patient workflows. You can use it either
as a full end-to-end pipeline or as a set of composable building blocks
for custom experiment orchestration.
Where to next? Pick the card that matches what you came to do.
Start here →
Install soma, run your first pipeline, and pick how you'll use it: pipeline API, building blocks, or CLI.
Compose building blocks →
Mix and match encoders, aggregators, and decoders for custom orchestration.
Pick a task →
Classification, regression, survival, segmentation, or detection — with the methods that fit each.
Reproduce a benchmark →
Curate, run, and tolerance-check a registered benchmark — the EVA patch suite or the OCELOT detection challenge.