Pre-trained models

Users can jump start the process of creating a deep-learning model by:

• applying one of the provided pre-trained models to their data using the plugin

• evaluating the segmentation results

• fine-tuning these models to better suit their specific datasets

Below is a table listing available pre-trained models available for download from the plugin through the Download dialog under the Help button. If you use our pre-trained models in your own research, please cite us; citation info included in the table. Thank you!

Available pre-trained models by the Allen Institute for Cell Science

Category	Name	MegaSeg_v1
Model Properties	Description	CNN-based structure-agnostic 3D segmentation model, trained on 2600 open 3D fluorescent microscopy images of cellular structures in human hiPSCs
	Architecture	nnUNet
	Loss function	Generalized dice focal loss
	Optimization	Adam
	# of Epochs	1000
	Stopping Criteria	Validation loss not improving for 100 epochs
	System Trained On	Nvidia A100
Dependencies	CytoDL Version	1.7.1
	PyTorch Version	2.4.0+cu118
Training Data	Image Resolution	55x624x924; 60x624x924; 65x600x900; 65x624x924; 70x624x924; 75x624x924; 75x600x900
	Microscope Objective	100x
	Microscopy Technique	Spinning disk confocal
	Public Data Link	Dataset Link
	Expected Performance	On NVIDIA-A100, 80GB, Inference @ 6.01 Secs for an Input image of size 924x624x65
	Structures Trained On	Actin bundles, ER(SERCA2), Adherens junctions, Desmosomes, Gap junctions, Myosin, Nuclear pores, Endosomes, ER (SEC61 Beta), Nuclear speckles, Golgi, Tight junctions, Mitochondria
Inference Data	Minimum Image Dimension	16x16x16
	Threshold value	We have validated MegaSeg using a 50% threshold or the threshold value of 128 i.e., pixels for which the model output is lower than 128 will be classified as background.
Disclosure	Info	Model will not be able to segment the tops and bottoms of nuclear(Lamin) and plasma(CAAX) membrane. Model will also face issue in segmenting dataset specific attributes (e.g., filled vesciles on lysosomes) in contrast to general attributes.
Citation	Info	The MegaSeg preprint is in preparation, if you use MegaSeg model in your own research, in the meantime you can cite Segmenter ML Plugin. Thank you!