rendeirolab/LazySlide

What it solves

LazySlide provides a unified Python framework for analyzing whole slide images (WSI) in digital and computational pathology. It simplifies the complex workflow of moving from raw slide images to high-level biological insights, automating tasks like tissue detection, tiling, and feature extraction using state-of-the-art AI models.

How it works

The framework integrates various tools and models to create a seamless pipeline. It uses wsidata to support a wide range of slide formats (SVS, NDPI, etc.) and stores results in the SpatialData format, ensuring interoperability with the scverse ecosystem (scanpy, anndata, squidpy). The pipeline typically involves:

• Preprocessing: Detecting tissue and tiling it at specific resolutions.
• Feature Extraction: Using pathology foundation models (like UNI or Virchow) or standard vision models to extract tile features.
• Analysis: Performing cell segmentation, zero-shot classification, and linking morphology to gene expression.

Who it’s for

It is designed for researchers and practitioners in digital pathology and computational biology who need an accessible way to integrate foundation models and spatial analysis tools into their research workflows.

Highlights

• Extensive Model Support: Access to 30+ pathology foundation models and vision-language models (CONCH, PLIP, TITAN) for zero-shot predictions and captioning.
• Interoperable Data: Native integration with the scverse ecosystem via SpatialData.
• Broad Format Support: Compatible with numerous slide formats including SVS, NDPI, MRXS, and DICOM.
• Deep Learning Ready: Provides PyTorch datasets for users to train their own custom models.