Atlas Toolkit: Fast registration of 3D morphological datasets in the absence of landmarks

Image registration is a gateway technology for Developmental Systems Biology, enabling computational analysis of related datasets within a shared coordinate system. Many registration tools rely on landmarks to ensure that datasets are correctly aligned; yet suitable landmarks are not present in many datasets. Atlas Toolkit is a Fiji/ImageJ plugin collection offering elastic group-wise registration of 3D morphological datasets, guided by segmentation of the interesting morphology. We demonstrate the method by combinatorial mapping of cell signalling events in the developing eyes of chick embryos, and use the integrated datasets to predictively enumerate Gene Regulatory Network states.

4 ºC to counter-stain cell nuclei. Reversing the sequence of Scale A2 > Propidium Iodide causes the latter to be washed from the specimen whereas antibody labelling remains intact.
Each specimen was mounted for microscopy by punching a hole through three or four layers of electrical tape before adhering to a clean glass microscope slide. A specimen was pipetted into the resulting chamber, ensuring that the level of AF1 mounting reagent protruded slightly above the surrounding electrical tape. A clean rectangular cover glass was then applied over the top, ensuring that air bubbles were excluded, and secured in place with nail polish. Once mounted, specimens were stored inverted at 4 ºC overnight.

Two-photon microscopy of whole-mount chick embryos.
Two-photon microscopy was performed using a La Vision Biotech TriMScope II instrument with inverted stand and ImSpector Pro software. Typically, a volume of 500 um x 500 um x 250 um was imaged using a 20X air objective with a numerical aperture of 0.8. AlexaFluor488 and Propidium Iodide underwent simultaneous two-photon excitation with a single laser line (Coherent Vision II Ti:Sapphire, pulsed femtosecond laser) at a wavelength of 930 nm and scan frequency of 200 Hz. AlexaFlour488 and Propidium Iodide fluorescence were separated using emission filters at 525 +/-25 nm and 620 +/-30 nm, respectively, and captured using a pair of sensitive, non-descanned GaAsP detectors. Two-photon image stacks were generated at an oversampled resolution of 0.333 µm (X) by 0.333 µm (Y) by 0.72 µm (Z). The resulting datasets were exported as OME Tiff files for subsequent analysis using the Fiji distribution of ImageJ 3 .

Relative quantification of nuclear protein levels in Fiji/ImageJ.
"Atlas Toolkit > 1. Extract Nuclear Signal" To account for variable depth penetration and shadowing across the complex tissue morphology of the 3D specimens, signalling protein levels indicated by AlexaFlour488 fluorescent signal was normalised to that of the Propidium Iodide nuclear counter stain.
Moreover, since the signalling proteins under investigation translocate the cell nucleus upon activation, the Propidium Iodide fluorescent signal was also used to isolate nuclear fluorescence as has been performed previously 4 .
This was necessary for Smad2, Smad3 and β-catenin proteins, for which the antibodies utilised do not report phosphorylation status. It is also necessary for the utilised phospho-Smad1/5/8 antibody since phosphatase treatment of specimens revealed that this antibody strongly labelled an unidentified and non-phosphorylated epitope in the cytoplasm of mitotic cells, in addition to the intended nuclear phosphorylated-Smad1/5/8 signal.
The Atlas Toolkit includes a tool "1. Extract Nuclear Signal" that recapitulates the standardised method used for quantifying nuclear AlexaFlour488 fluorescence relative to that of Propidium Iodide in this study. This method may need to be adjusted for different biological specimens and imaging conditions.
The method for relative quantification is schematised in Figure 2b. The Propidium Iodide signal (channel 1) was subjected to Auto-Local Thresholding in Fiji/ImageJ in order to generate a binary mask were the locations of cell nuclei were represented with a pixel value of 255 (white), where as non-nuclear pixels had a value of zero (black '.label' files are used both to project nuclear signalling activities onto the optic vesicle tissue morphology, and to perform non-rigid registration of data sets from different biological specimens. As an aid to identifying closely associated tissue layers, it is useful to merge both the Propidium Iodide and AlexaFluor488 data channels, and particularly to increase the brightness of the AlexaFluor488 channel since the faint background auto-fluorescence of this channel can help to distinguish tissue boundaries. Moreover, it is also helpful to reduce the X and Y resolution of the merged image stack (known as down-sampling) by a factor of three

Projection of signalling levels onto optic vesicle tissue morphology in Fiji/ImageJ.
"Atlas Toolkit > 2. Project to Segment Label" In order to relate molecular patterning events with tissue morphology, it is necessary to be able to 'project' the former onto the latter. The Atlas Toolkit includes a plugin "2. Project to Segment Label", which automates this process and makes use of both the normalised fluorescent signal (Step 3) and a segmented '.label' file (Step 4). This plugin works by firstly isolating the tissue of interest from the normalised data stack (according to the supplied '.label' file, allowing the user to select from multiple labels when present). It then divides the isolated data segment into a voxel lattice of a specified size (the 'Voxel Size' parameter measured in pixels or µm depending on the calibration of the normalised data stack; 12 µm was used for this study). Each voxel is then assigned a value corresponding to the mean average signal level of its local neighbourhood (the 'Sample Radius' parameter measured in voxels; a radius of three voxels was used for this study). Non-nuclear pixels (which have a zero value) are excluded from averaging, but are assigned the same value as nuclear pixels within the same voxel. This local averaging smoothens local fluctuations and signal noise, and fills the segmented tissue with the locally averaged signal level. The resulting cubic lattice is then 'cropped' to the correct surface morphology using the supplied '.label' file.

Group-wise elastic registration of 3D segmented data sets in Fiji/ImageJ.
"Atlas Toolkit > 3. Label Registration 3D" "Atlas Toolkit > 4. Apply Label Registration" "Atlas Toolkit > 5. Merge Registered Volumes" In order to compare molecular patterning or other quantities between independent specimens, it is necessary to first align (register) the 3D segmented tissues. This task is particularly difficult due to natural variation in tissue morphology at microscopic scales, difficulties in precise stage matching of specimens, and variation in specimen orientation Cluster3.0, as well as a 'consensus.tif' file written by the "3. Label Registration 3D" plugin (Step 6, above). In addition to the 3D reconstruction, it generates a dendrogram representing a user-defined subset of the hierarchical tree (only the longest branches are shown in the present study), and a 2D heat-map representation of the mean quantities (e.g. nuclear protein levels in this study) within each cluster.

Numerical evaluation of Atlas Toolkit registration performance using a synthetic dataset.
Since real optic vesicle tissues lack objective landmarks, it was necessary to evaluate registration performance using a simulated dataset that was modified to include synthetic landmarks.
The segment label from a real Optic Vesicle object was augmented by addition of seven arbitrarily positioned spherical landmarks. The landmarks and segment label where then separated into two files, and both were identically deformed by translation, rotation and/or scaling, in order to create three manually deformed datasets. The three manually deformed segment labels were registered together using our tool "3. Label Registration 3D" for between one and six iterations. The resulting '.ots' files were then used to transform the three corresponding landmark files into the shared coordinate system of the consensus object. Following landmark transformation, the Fiji/ImageJ 3D Object Counter function was used to locate the centres of the spherical landmarks, and the pair-wise distances between corresponding landmarks were calculated (Fig. 1f -h; Supp. Fig. S1a). The mean pair-wise distance between all corresponding landmark pairs following one iteration was 6.12 µm +/-4.78 µm (mean +/-standard deviation), or 3.24 µm +/-3.6 µm (mean +/-standard deviation) following six iterations.
In addition to landmark distances, we also measured the mean volumetric overlap between all three optic vesicles and their shared consensus morphology ( Supplementary Fig.   S1b). For one iteration, the mean volumetric overlap was 93.26 % +/-0.19 % (mean +/standard deviation), and for six iterations it was 95.85 % +/-0.14 % (mean +/-standard deviation).
The time required to compute these registrations was 180 seconds (for one iteration and 1,680 seconds for six iterations (Supp. Fig. S1c), on a MacBook Pro laptop computer with 2.7 GHz Intel Core i7 CPU and 16 GB of RAM, running on battery power.

Numerical comparison of Atlas Toolkit with two BrainAligner methods.
In order to better gauge the performance of our landmark-free method, we used the same  Fig. S1c). In this situation BrainAligner is almost 10 times less accurate than Atlas Toolkit. This is possibly because the 3D datasets undergoing registration contain purely binary information: pixel values indicate the presence or absence of tissue at each position in the volume. It is likely there is insufficient information for BrainAligner to accurately identify corresponding landmark features between the different 3D datasets.

Numerical comparison of Atlas Toolkit with the Computational Morphometry
Toolkit (CMTK).
This is in contrast to Atlas Toolkit, which decomposes the 3D registration problem into a series of orthogonal 2D elastic registrations. Consequently, 'groupwise_warp' might be expected to generate more accurate registrations at the cost of greatly increased computation time. We therefore tested it on the same synthetic dataset used to evaluate Atlas Toolkit and BrainAligner.

Known Limitations
We have used Atlas Toolkit to successfully register various 'thick' epithelial tissue morphologies including optic vesicle tissues from HH10 -HH14, and optic cup (Supp. Fig.   S2a -c) and lens vesicle tissues (Supp. Fig. S2d -f) Fig. S2g & h), the thin and undulating morphologies of squamous epithelia often fail to achieve good intersections (e.g. HH16 surface ectoderm, Supp. Fig. S2d -f; schematised in Supp. Fig. S2i & j). This failure to register squamous epithelia appears to arise as the tissue's smallest dimension approaches the magnitude of registration error produced by Atlas Toolkit (Fig. 1f; Supp. Fig. S1a). The test data included with Supplementary Software 1 provides examples of both success (HH10 optic vesicle; label channel 1) and fail cases (HH10 surface ectoderm; label channel 2) within the same datasets.
Manual segmentation of entire objects from raw datasets is more time-consuming than placing a small number of precise fiduciary markers. It will also likely require training of personnel to accurately distinguish closely associated tissues (e.g. surface ectoderm versus optic vesicle epithelia in the developing eye) in the absence of molecular labels, and especially where optical sectioning is oblique to the junction between tissues. However, fiduciary marker placement depends upon the presence of user-identifiable landmarks, but not all tissues are rich in small landmark features (e.g. the optic vesicle, limb bud, heart tube).
Thus, segmentation-guided registration as provided by Atlas Toolkit may be preferable for landmark-poor tissues.