Robust Non-rigid Registration Challenge for Expansion Microscopy Banner

Dataset

Sample Preparation and Imaging Process

Expansion microscopy (ExM) is an imaging technique where a specimen is embedded in a swellable polymer mesh and is physically expanded, allowing the conventional fluorescence microscope to resolve features of interest at a nanoscale level [1]. 

In particular, the use of ExM in conjunction with multiplexed imaging is a promising approach to simultaneously studying a multitude of target objects (e.g. nucleobases and proteins) in intact tissues. To achieve this, multiple rounds of imaging are performed for the same expanded sample with different sets of staining to yield a collection of multi-round, multi-channel, high-resolution 3D image volumes capturing one target object per round per channel. For instance, in transcriptomics, a field of study that has led to developments in personalized medicine, drug discovery, and cancer research, it has been shown that multiplexed ExM imaging for nucleobases of RNAs can be used to investigate spatially-resolved transcriptomics, where RNAs can be localized even in dendritic spines of neurons [2].

To analyze a sample with multi-round staining, we need registration methods to align 3D multi-channel image volumes across imaging rounds because the gel-like ExM sample deforms/the microscope moves while applying a new set of staining for each subsequent round. Therefore, a typical solution is to image the same target in one of the channels repeatedly for all the rounds and use it as the anchoring channel to compute deformation between rounds.

Because ExM super-resolution imaging requires precise registration, it is essential to have methods that can handle various non-idealities: movement of the microscope, repetitive small-scale texture, degraded image quality over rounds due to re-staining, and deformation of gel. Thus, we're releasing data from three commonly imaged species to facilitate the development of a robust registration pipeline for ExM registration: zebrafish (brain), mouse (cortex), and C. elegans (whole).

Training Dataset

The training dataset is now available on Zenodo. Each species highlights a different challenge of ExM registration: 

**Mouse: **

  • Has non-rigid deformation of the hydrogel and loss of staining intensity. Deformation of the hydrogel occurs because the sample sits for multiple days and at a low temperature between staining rounds.
  • One volume pair in particular has an extreme difference in the strength of the signal, making registration difficult.
  • There are four training volume pairs. 
  • The voxel spacing is [0.1625 um, 0.1625um, 0.4 um] for both the fixed and moving volumes.

Zebrafish:

  • Has a strong z-offset, repetitive small-scale textures (circular nuclei), and mostly rigid deformations. 
  • There are four training volume pairs. 
  • The voxel spacing is [0.1625 um, 0.1625um, 0.4 um] for both the fixed and moving volumes.

**C. elegans: **

  • Has a strong z-offset, large volume size (300-500+ z-slices), and mostly rigid deformations. 
  • There are four training volume pairs. 
  • The voxel spacing is [0.1625 um, 0.1625um, 0.4 um] for both the fixed and moving volumes.

The data is saved in an .h5 format. Within each .h5 file is a pair of volumes: a fixed volume and a volume to align. You can open the images using the Python code here. 

When creating models, please be mindful of the runtime (we are particularly interested in solutions that are fast!). In a real-world setting, a single experiment may need hundreds of volume pairs aligned. **For this reason, we suggest experimenting with learning-based approaches that can quickly predict deformation fields once trained and can be fine-tuned for new experiments. **

Want more training data? Use our code to generate additional rigid and non-rigid synthetic deformations. 

Validation Dataset

The validation dataset is now available on Zenodo. Participants can also view the evaluation algorithm. There is one validation volume pair per species. All have a spacing of [0.1625 um, 0.1625um, 0.4 um].

Test Dataset

The test dataset and example .h5 submission files are now available on Zenodo.  Segmentation maps are hidden from participants to prevent overfitting to the sparse landmarks. There are three pairs of volumes per species, all with a spacing of [0.1625 um, 0.1625um, 0.4 um]. Participants are asked to submit dense deformation fields for each of the volume pairs in the test dataset. For information on the submission format, please refer to the challenge rules.  Code to assist in generating the .h5 files has been added to the GitHub. The leaderboards will open on 2/15 and will allow for one submission per day.

[1] F. Chen, P. W. Tillberg, and E. S. Boyden, “Expansion microscopy,” Science, vol. 347, no. 6221, pp. 543–548, 2015.

[2] S. Alon, D. R. Goodwin, A. Sinha, A. T. Wassie, F. Chen, E. R. Daugharthy, Y. Bando, A. Kajita, A. G. Xue, K. Marrett et al.,  “Expansion sequencing: Spatially precise in situ transcriptomics in intact biological systems,” Science, vol. 371, no. 6528, p. eaax2656, 2021.