bism - Biomedical Image Segmentation Models

BISM is a repository for training and evaluating biomedical instance segmentation models -- something akin to the timm package for 2D image tasks, but 3D instance segmentation. When at all possible, each model will offer a 2D or 3D implementation, however we will not provide pre-trained model files.

No Documentation right now. In general, you launch a training run through a yaml configuration file. Check out bism.train.__main__.py as the starting point for training. bism.config.config.py for the default configuration for each approach. This should (hopefully) allow for repeatable training of 3D instance segmentation models of various types.

To execute a training config, simply run python bism/train --config_file "Path/To/Your/File.yaml". To run a pretrained model, simply run python bism/eval -m "path/to/model/file.trch" -i "path/to/image.tif" To launch the model inspector, run python bism/gui

This module is under active development so should not be used for anything but research purposes!

Current Models

Model	2D	3D	Scriptable
UNet	✓	✓	✓
UNeXT	✓	✓	✓
Recurrent UNet	✓	✓	✓
Residual UNet
Unet++	✓	✓	✓
CellposeNet	✓	✓	✓

Current Generic Blocks

BLOCK NAME	2D	3D
UNeXT Block	✓	✓
ConcatConv	✓	✓
Recurrent UNet BLock	✓	✓
Residual UNet BLock	✓	✓
DropPath	✓	✓
LayerNorm	✓	✓
UpSample	✓	✓
ViT Block

Segmentation Implementation

APPROACH	2D	3D
Cellpose
Affinities		✓
Local Shape Desc.		✓
Omnipose		✓
Auto Context LSD		✓
Multitask LSDs		✓
Semantic	✓	✓
Mask RCNN	✓

Loss Functions

Function	Implemented
Dice	✓
CL Dice	✓
Tverksy	✓
Jaccard	✓