Choose a starting model
Use a task-specific model when its task and classes fit your study. A starter model instead supplies initial weights for learning a new task from your annotations. Install the module containing your chosen model, then select it in Train. You do not need to edit its recipe.
Choose by the objects you need
Each method has two starting-weight options, explained below. The links open the starter modules’ model cards on Hugging Face.
Method |
Start here for |
H&E-12 weights |
DINOv3 weights |
|---|---|---|---|
Cellpose |
Nuclei, cells and irregular or elongated objects |
||
HoVer |
Large individual tissue structures |
||
Dense semantic |
Tissue compartments, rather than separate cells |
Both Cellpose and HoVer can separate touching objects and segment nuclei. These are starting recommendations, not a guarantee that one will be more accurate or faster on your images. Compare the trained result on held-out images, including crowded areas and unusual shapes.
Cellpose groups pixels into objects by following predicted flows. HoVer uses watershed separation, without repeated flow-following steps. Cellpose can also segment large structures, but objects spanning many pixels may need more flow-following iterations to form complete objects. These are steps during prediction, not training epochs. HistoForge fixes this iteration count internally; increasing training epochs does not change it. Object size here means its size in pixels at the chosen resolution, not the overall slide size.
Choose the starting weights
Each method has two starting-weight families:
Family |
Starting point |
|---|---|
H&E-12 |
Histology-pretrained weights. The Cellpose starter retains its pretrained segmentation model; the HoVer and semantic variants have newly initialized prediction layers. |
DINOv3 |
A pretrained image encoder, with newly initialized layers for histology segmentation. |
Starter labels such as Object and Foreground are generic, not learned biological classes.
Define your target classes and resolution during training. Newly initialized prediction layers
need training before their results are meaningful; pretrained components do not establish
accuracy on your images. Review the starter’s model card for its initialization and terms.
The Train tab lists only installed modules containing a model supported for training by the connected backend. Within a module, only supported models can be selected. A model can be available for analysis without having a supported training implementation.
Install your chosen module, then prepare your training data.