Requirements

This section outlines the software and system requirements needed to run AutoMIL.

Platform / Operating System

AutoMIL builds on top of the Slideflow framework for WSI processing, dataset management, and model training. While AutoMIL itself is platform-agnostic, slideflow is primarily developed and tested on Linux and depends on system-level libraries such as cuCIM. cuCIM is part of the RAPIDS collection of GPU-accelerated software solutions for data science, all of which are developed for usage on Linux. Consequently, full functionality, stability, and performance of AutoMIL can only be guaranteed on Linux, which is therefore strongly recommended.

Python Version

AutoMIL requires Python ≥ 3.11. A 64-bit Python installation is strongly recommended, especially when working with large whole-slide images.

Core Python Dependencies

The following Python packages are required and are installed automatically when installing AutoMIL via pip install .:

Deep Learning and Numerical Computing

torch
torchvision
numpy
pandas
scikit-learn
fastai

Visualization and Interactive Analysis

matplotlib
seaborn
imgui
ipython

Whole-Slide Image Processing and MIL Utilities

openslide-python
openslide-bin
slideflow
nystrom-attention
cucim

General Utilities and Command Line Interface

click
pyrfr
swig

No manual installation of these dependencies is required when installing AutoMIL from source.

Optional Dependencies

`pyvips`/`libvips`

By default, Slideflow uses the image processing library cuCIM for handling WSIs. In certain edge cases, however, cuCIM is not a reliable solution for image processing tasks. To our knowledge, these cases include:

Working with OME-TIFF files
OME-TIFF is a common WSI format that combines TIFF image data with XML metadata. As of now, OME-TIFF is not supported by cuCIM.
Using AutoMIL’s PNG → TIFF conversion pipeline
For PNG-based datasets (which are generally not recommended for WSIs, but are sometimes used in practice), AutoMIL provides an opt-in preprocessing step to convert PNG images to TIFF. cuCIM is not well suited for processing very large PNG images in this context.

For both cases, AutoMIL provides an optional dependency group that installs the pyvips library as an alternative image processing backend. Install it via:

pip install .[vips]

System dependency required for pyvips

pyvips is only the Python binding for the image processing library libvips, which must be installed separately on your system.

sudo apt install libvips