Available Models

Denoising Dense Encodings Models

Encodings presented in our manuscript suppress visual artifacts in EM images while highlighting biological structures. These encodings are helpful both for alignment and for training auxilary models, such as fold, crack and plastic detection models.

We provide several encoder models listed below. Encoder models expect normalized image data as input and generalize well accross species and datasets. Each encoder outputs single channel floating point tensor. Note that encoding value ranges may differ across scales.

Encoder Models

Path	XY Resolution Range
gs://corgie_package/models/encoders/encoder_16_64nm	16-64nm
gs://corgie_package/models/encoders/encoder_64_256nm	64-256nm
gs://corgie_package/models/encoders/encoder_256_2048nm	256-2048nm

Example usage command for an encoder is as follows:

corgie apply-processor \
--src_layer_spec '{"path": "'${CORGIE_WALKTHROUGH_PATH}'/img/img_norm"}' \
--dst_layer_spec '{
   "type": "img",
   "path": "'${CORGIE_WALKTHROUGH_PATH}'/mip7_enc",
   "args": {"dtype": "float32"}
   }' \
--processor_spec '{
    "ApplyModel": {
        "params": {"path": "gs://corgie_package/models/encoders/encoder_256_2048nm"}
    }
}' \
--processor_mip 7 \
--device "cuda" \
--start_coord "100000, 100000, 17000" \
--end_coord "150000, 150000, 17001" \
--verbose \
--chunk_xy 1024 --chunk_z 1

Aligners

Online Finetuner

Online Finetuner uses gradient descent to optimize the alignment field as described in our manuscript. Online Fintetunnr can practically correct misalignments up to 2px at the resolution it’s being applied at, and reduce the remaining misalignment to below 0.5px.

Our Online Finetunner alignment model is publically available for use by specifying processor path to gs://corgie_package/models/encoders/online_finetunner. The model operates on float32 normalized images or encodings as input. Rigidity penalty weight is specified by the sm=200 constructor argument. Setting sm to values smaller than 200 will result in less rigid tissue deformations and vice verca. The number of iterations used by the Online Finetuner is determined by the num_iter==2000 parameter. Given a good rigidity penalty weight value, Online Fintetuner has been shown to generalize well accross species and datasets.

MICrONS Coarse Aligner

MICrONS dataset contains particularyly challenging folds of large magnitude. We release the model that was used to correct these folds, which can be used by setting the path to gs://corgie_package/models/aligners/MICrONS_aligner_512_1024nm.