The pipeline
The ideas
Foundation model. Our models are pre-trained on enormous, diverse histology data, so they produce rich, general-purpose representations that transfer to many tasks without task-specific training from scratch. Why tiling. A whole slide image can be billions of pixels — far too large to process at once. The slide is split into small tiles, each processed independently. Tissue masking first. Slides are mostly background. The pipeline first tiles the slide coarsely (512×512 at 8 µm/px) and runs tissue segmentation to produce a tissue map. Only tissue-bearing regions are then re-tiled finely (224×224 at 0.5 µm/px) for the model — cutting cost and time before the expensive feature step. Embeddings (H-Optimus). For each tile, H-Optimus outputs a 1536-d feature vector — a compact numerical fingerprint of what the tissue looks like. You feed these into your own models for prediction, retrieval, or classification. Prediction + embeddings (M-Optimus). M-Optimus produces the same 1536-d tile embeddings and goes further: it predicts spatial gene expression directly from the tile (optionally informed by bulk RNA), giving a molecular readout without a spatial assay. One pipeline, multiple outputs. The SDK wires this together — point it at a slide and a model, and it tiles, masks, dispatches, and writes results to disk.Pick the right model
H-Optimus vs. M-Optimus, side by side.