pathml

Whole-slide imaging formats like SVS and NDPI are converted into structured data using unified WSI loaders that extract tiles, pyramids, and metadata for machine learning preprocessing. This process transforms raw pathology slides into clean computational inputs for segmentation and modeling.

A reproducible H&E stain normalization and nucleus detection pipeline is built by composing preprocessing transforms for tissue detection, stain normalization, and mask generation. These composable pipelines ensure consistent preprocessing outputs for downstream analysis.

Multiparametric CODEX, Vectra, and MERFISH imaging can be processed for spatial graph analysis by collapsing multiplex runs and segmenting cells. Marker expression is quantified into AnnData to support configurable spatial omics graph construction with knn, radius, or delaunay connectivity.

The pathology ML workflow supports PyTorch-integrated training and prebuilt nucleus models like HoVer-Net for tile-based inference. It also includes ONNX deployment support to operationalize trained segmentation models within computational pathology pipelines.

Large pathology tiles and masks are managed for scalable batch processing using HDF5-backed storage. This approach persists tiles, masks, features, and metadata to enable fast random access and reproducible analytics across extensive whole-slide imaging datasets.

Spatial graphs for cell and tissue segmentation work by converting segmentation masks into graph structures with configurable connectivity like knn, radius, or contact. These graphs include rich node and edge features for spatial omics modeling and downstream machine learning analysis.