Atlas Interpolation

The Allen Brain Institute hosts a rich database of mouse brain imagery. It contains a large number of gene expression datasets obtained through the in situ hybridization (ISH) staining. While for a given gene a number of datasets corresponding to different specimen can be found, each of these datasets only contains sparse section images that do not form a continuous volume. This package explores techniques that allow to interpolate the missing slices and thus reconstruct whole gene expression volumes.

• Installation
  • Python Version and Environment
  • Install "Atlas Interpolation"
• Data
  • Remote Storage Access
  • Model Checkpoints
  • Section Images and Datasets
  • New ISH datasets (advanced, optional)
• Examples
  • Pair Interpolation
  • Optical Flow Models
  • Predict an Entire Gene Volume (longer runtime)
• Vendors
• Funding & Acknowledgment

Installation

Python Version and Environment

Note that due to some of our dependencies we're currently limited to python version 3.7. Please make sure you set up a virtual environment with that version before trying to install this library. If you're unsure how to do that please have a look at conda or pyenv.

If you are part of the Blue Brain Project and are working on the BB5 you can find the correct python version in the archive modules between archive/2020-02 and archive/2020-12 (inclusive). Here's an example of a set of commands that will set up your environment on the BB5:

module purge
module load archive/2020-12
module load python
python -m venv venv
. ./venv/bin/activate
python --version

We also recommend that you make sure that pip is up-to-date and that the packages wheel and setuptools are installed:

pip install --upgrade pip wheel setuptools

Install "Atlas Interpolation"

In order to access the data and the example scripts a local clone of this repository is required. Run these commands to get it:

git clone https://github.com/BlueBrain/atlas-interpolation
cd atlas-interpolation

The "Atlas Interpolation" package can now be installed directly from the clone we just created:

pip install '.[data, optical]'

Data

The data for this project is managed by the DVC tool and all related files are located in the data directory. The DVC tool has already been installed together with the "Atlas Interpolation" package. Every time you need to run a DVC command (dvc ...) make sure to change to the data directory first (cd data).

Remote Storage Access

We have already prepared all the data, but it is located on a remote storage that is only accessible to people within the Blue Brain Project who have access permissions to project proj101. If you're unsure you can test your permissions with the following command:

ssh bbpv1.bbp.epfl.ch \
"ls /gpfs/bbp.cscs.ch/data/project/proj101/dvc_remotes"

Possible outcomes:

# Access OK
atlas_annotation
atlas_interpolation

# Access denied
ls: cannot open directory [...]: Permission denied

Depending on whether you have access to the remote storage in the following sections you will either pull the data from the remote (dvc pull) or download the input data manually and re-run the data processing pipelines to reproduce the output data (dvc repro).

If you work on the BB5 and have access to the remote storage then run the following command to short-circuit the remote access (because the remote is located on the BB5 itself):

cd data
dvc remote add --local gpfs_proj101 \
  /gpfs/bbp.cscs.ch/data/project/proj101/dvc_remotes/atlas_interpolation
cd ..

Model Checkpoints

Much of the functionality of "Atlas Interpolation" relies on pre-trained deep learning models. The model checkpoints that need to be loaded are part of the data.

If you have access to the remote storage (see above) you can pull all model checkpoints from the remote:

cd data
dvc pull checkpoints/rife.dvc
dvc pull checkpoints/cain.dvc
dvc pull checkpoints/maskflownet.params.dvc
dvc pull checkpoints/RAFT.dvc
cd ..

If you don't have access to the remote you need to download the checkpoint files by hand and put the downloaded data into the data/checkpoints folder. You may not need all the checkpoints depending on the examples you want to run. Here are the instructions for the four models we use: RIFE, CAIN, MaskFlowNet, and RAFT:

• RIFE: download the checkpoint from a shared Google Drive folder by following this link. Unzip the contents of the downloaded file into data/checkpoints/rife. [ref]
• CAIN: download the checkpoint from a shared Dropbox folder by following this link. Move the downloaded file to data/checkpoints/cain. [ref]
• MaskFlowNet: download the checkpoint directly from GitHub by following this link. Rename the file to maskflownet.params and move it to data/checkpoints. [ref]
• RAFT: download the checkpoint files from a shared Dropbox folder by following this link. Move all downloaded .pth files to the data/checkpoints/RAFT/models folder. [ref]

If you downloaded all checkpoints or pulled them from the remote you should have the following files:

data
└── checkpoints
    ├── RAFT
    │   ├── models
    │   │   ├── raft-chairs.pth
    │   │   ├── raft-kitti.pth
    │   │   ├── raft-sintel.pth
    │   │   ├── raft-small.pth
    │   │   └── raft-things.pth
    ├── cain
    │   └── pretrained_cain.pth
    ├── maskflownet.params
    └── rife
        ├── contextnet.pkl
        ├── flownet.pkl
        └── unet.pkl

Section Images and Datasets

The purpose of the "Atlas Interpolation" package is to interpolate missing section images within section image datasets. This section explains how to obtain these data.

Remember that if you don't have access to the remote storage (see above) you'll need to use the dvc repro commands that download/process the data live. If you do have access, you'll use dvc pull instead, which is faster.

Normally it's not necessary to get all data. Due to its size it may take a lot of disk space as well as time to download and pre-process. If you still decide to do so you can by running dvc repro or dvc pull without any parameters.

Specific examples only require specific data. You can use DVC to list all data pipeline stages to find out which stage produces the data you're interested in. To list all data pipeline stages run:

cd data
dvc stage list

If, for example, you need data located in data/aligned/coronal/Gad1, then according to the output of command above the relevant stage is named align@Gad1. Therefore, you only need to run this stage to get the necessary data (replace repro by pull if you can access the remote storage):

dvc repro align@Gad1

New ISH datasets (advanced, optional)

If you're familiar with the AIBS data that we're using and would like to add new ISH gene expressions that are not yet available as one of our pipeline stages (check the output of dvc stage list) then follow the following instructions.

1. Edit the file data/dvc.yaml and add the new gene name to the lists in the stages:download_dataset:foreach and stages:align:foreach sections.
2. Run the data downloading and processing pipelines (replace NEW_GENE by the real gene name that you used in data/dvc.yaml):

   dvc repro download_dataset@NEW_GENE
   dvc repro align@NEW_GENE
   

Examples

In this section we showcase several typical use-cases of "Atlas Interpolation":

• Use pair interpolation to predict an intermediate image between two given images
• Predict optical flow between any pair of images and use it to morph a third image
• In a gene expression volume predict missing slices and reconstruct the whole volume

Note that all models accept both RGB images (shape=(height, width, 3)) and grayscale images (shape=(height, width)).

Pair Interpolation

The only data you need for this example is the RIFE model checkpoint. Follow the instructions in the corresponding section above to get it. If you have access to the remote data storage it's enough to run the following commands:

cd data
dvc pull checkpoints/rife.dvc
cd ..

In this example we start with a pair of images img1 and img2 (randomly generated for example's sake). First use the RIFE model to interpolate between them in a manual way and find the image in-between (img_middle). Then we demonstrate the use of the PairInterpolate class that streamlines the interpolation procedure. Starting with the same pair of images we iterate the interpolation three times to produce a stack of seven interpolated images (interpolated_imgs).

import numpy as np

from atlinter.vendor.rife.RIFE_HD import Model as RifeModel
from atlinter.vendor.rife.RIFE_HD import device as rife_device
from atlinter.pair_interpolation import PairInterpolate, RIFEPairInterpolationModel

# Get the input images
img1 = np.random.rand(100, 200, 3) # replace by real section image
img2 = np.random.rand(100, 200, 3) # replace by real section image

# Get the RIFE interpolation model
checkpoint_path = "data/checkpoints/rife/" # Please change, if needed
rife_model = RifeModel()
rife_model.load_model(checkpoint_path, -1)
rife_model.eval()
interpolation_model = RIFEPairInterpolationModel(rife_model, rife_device)

# Manually predict middle image between img1 and img2
preimg1, preimg2 = interpolation_model.before_interpolation(img1=img1, img2=img2)
img_middle = interpolation_model.interpolate(img1=preimg1, img2=preimg2)
img_middle = interpolation_model.after_interpolation(img_middle)
print(img_middle.shape)

# Streamline the interpolation using PairInterpolate and predict a stack
# of 7 intermediate images
interpolator = PairInterpolate(n_repeat=3)
interpolated_imgs = interpolator(img1, img2, interpolation_model)
print(interpolated_imgs.shape)

Optical Flow Models

The only data you need for this example is the MaskFlowNet model checkpoint. Follow the instructions in the corresponding section above to get it. If you have access to the remote data storage it's enough to run the following commands:

cd data
dvc pull checkpoints/maskflownet.params.dvc
cd ..

This example demonstrates how an optical flow model can be used to compute the optical flow between a pair of images. It can then be used to warp a third image. The images in this example are randomly generated. In a realistic setting they should be replaced by real images.

import numpy as np

from atlinter.optical_flow import MaskFlowNet

# Instantiate an optical flow model (in this case: MaskFlowNet)
checkpoint_path = "data/checkpoints/maskflownet.params"
net = MaskFlowNet(checkpoint_path)

# Prepare random images. Should be replaced by real section images
img1 = np.random.rand(100, 200, 3)
img2 = np.random.rand(100, 200, 3)
img3 = np.random.rand(100, 200, 3)

# Predict the optical flow between img1 and img2
img1, img2 = net.preprocess_images(img1=img1, img2=img2)
predicted_flow = net.predict_flow(img1=img1, img2=img2)

# Warp a third image using the optical flow
predicted_img = net.warp_image(predicted_flow, img3)
print(predicted_img.shape)

Predict an Entire Gene Volume (Longer Runtime)

The data you need for this example are the RIFE model checkpoint and the Vip gene expression dataset. To get the RIFE checkpoint follow the instruction in the corresponding section above. If you have access to the remote data storage it's enough to run the following commands:

cd data
dvc pull checkpoints/rife.dvc
cd ..

As described in the data section above, there are two ways of getting the Vip gene expression dataset. If you have access to the remote data storage you can pull it from there:

cd data
dvc pull download_dataset@Vip
cd ..

If you don't have access then you can re-download it. This should always work, but may take several minutes:

cd data
dvc repro download_dataset@Vip
cd ..

In this example with start with a gene expression volume that has missing section images. First we load the image data and the metadata from disk and wrap it into a GeneDataset class. Then we instantiate the RIFE deep learning model that will be used for interpolation. We use this model to first predict a single slice in the volume, then we reconstruct the whole volume by predicting all intermediate slices. Note that this last step is computation-heavy and might therefore take some time.

import json

import numpy as np

from atlinter.data import GeneDataset
from atlinter.pair_interpolation import GeneInterpolate, RIFEPairInterpolationModel
from atlinter.vendor.rife.RIFE_HD import Model as RifeModel
from atlinter.vendor.rife.RIFE_HD import device as rife_device

# Load the gene expression dataset from disk
data_path = "data/sagittal/Vip/1102.npy"  # Change the path if needed
data_json = "data/sagittal/Vip/1102.json" # Change the path if needed
section_images = np.load(data_path)
with open(data_json) as fh:
    metadata = json.load(fh)

section_numbers = [int(s) for s in metadata["section_numbers"]]
axis = metadata["axis"]

# Wrap the data into a GeneDataset class
gene_dataset = GeneDataset(
  section_images,
  section_numbers,
  volume_shape=(528, 320, 456, 3),
  axis=axis,
)

# Load the RIFE deep learning model that will be used for interpolation
checkpoint_path = "data/checkpoints/rife"
rife_model = RifeModel()
rife_model.load_model(checkpoint_path, -1)
rife_model.eval()
rife_interpolation_model = RIFEPairInterpolationModel(rife_model, rife_device)

# Create a gene interpolator
gene_interpolate = GeneInterpolate(gene_dataset, rife_interpolation_model)

# Predict a single section image
predicted_slice = gene_interpolate.predict_slice(10)
print(predicted_slice.shape)

# Reconstruct the whole volume. This might take some time.
predicted_volume = gene_interpolate.predict_volume()
print(predicted_volume.shape)

Vendors

Some dependencies are not available as packages and therefore had to be vendored. The vendoring is done using the py-vendor utility. It's installed automatically together with the dev extras. You can also install it by hand via pip install py-vendor==0.1.2.

The vendoring is then done using the following command (add --force to overwrite existing folders):

py-vendor run --config py-vendor.yaml

See the py-vendor.yaml file for details on the vendor sources and files.

Funding & Acknowledgment

The development of this software was supported by funding to the Blue Brain Project, a research center of the École polytechnique fédérale de Lausanne (EPFL), from the Swiss government’s ETH Board of the Swiss Federal Institutes of Technology.

Copyright (c) 2021 Blue Brain Project/EPFL