Most computational pathology tools are validated for one clinical use case. This is the right approach for diagnostics. It's wrong for research. Research questions change project to project, and generalized tools fail to perform at an acceptable standard.
Trained and validated against the tissue types, stains, and questions each project requires, drawing from whichever of these fits, never a fixed general-purpose pipeline. Several of these approaches are what let a model detect virtual biomarkers, features with no established stain or scoring protocol, directly from morphology.
ABMIL and CLAM-style architectures for weakly-labeled whole-slide classification, where only slide-level ground truth exists.
Self-supervised, pathology-pretrained backbones fine-tuned or probed for the task, rather than trained from scratch on limited data.
Cell and tissue-neighborhood graphs that encode spatial context directly, for questions where position relative to neighbors is the signal.
Traditional segmentation and feature-engineering approaches where they outperform learned methods, on limited data or well-defined morphology.
Each one centers on a virtual biomarker: a feature no existing protocol could score, detected directly from morphology.
Replaced a 5-plex mIF + FISH workflow with one morphology-based model and a single serial-section read, for renal tubule identification and therapeutic agent localization. Read the full approach →
Early-stage: extending that same approach to liver, where autofluorescence makes mIF panels difficult and cell typing laborious. Read the full approach →