Deep Learning Based Focus Interpolation for Whole Slide Images

Model

The model is based on the U-Net architecture and by default consists of two down-sampling and up-sampling blocks.

The model accepts two different z-level images in the input and outputs:

• the intermediate z-level image (target-learning)
• two residual images that represent the difference between the target image and the input images (residual-learning)

Data

The model has been tested on the the image set BBBC006v1 from the Broad Bioimage Benchmark Collection Ljosa et al., Nature Methods, 2012.

The dataset is available at: https://bbbc.broadinstitute.org/BBBC006

Training

1. Download the BBBC006v1 dataset and prepare it according to the following structure:

   BBBC006_v1
   - train
   -- BBBC006_v1_images_z_00
   -- ...
   -- BBBC006_v1_images_z_33
   --- mcf-z-stacks-03212011_p23_s2_w2d36e0477-6528-4f5c-a76a-d76d199e07ca.tif
   --- ...
   - test
   ...
   

   and specify the dataset location in dfi/bbbc006_v1.py. Alternatively, implement a new method based on dfi/bbbc006_v1.py for a different dataset.

2. Specify the relevant paths in hparams.yaml

3. Execute $./run.sh