Constructing and Optimizing 3D Atlases From 2D Data With Application to the Developing Mouse Brain
This article has been Reviewed by the following groups
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
- Evaluated articles (eLife)
Abstract
3D imaging data necessitate 3D reference atlases for accurate quantitative interpretation. Existing computational methods to generate 3D atlases from 2D-derived atlases result in extensive artifacts, while manual curation approaches are labor-intensive. We present a computational approach for 3D atlas construction that substantially reduces artifacts by identifying anatomical boundaries in the underlying imaging data and using these to guide 3D transformation. Anatomical boundaries also allow extension of atlases to complete edge regions. Applying these methods to the eight developmental stages in the Allen Developing Mouse Brain Atlas (ADMBA) led to more comprehensive and accurate atlases. We generated imaging data from fifteen whole mouse brains to validate atlas performance and observed qualitative and quantitative improvement (37% greater alignment between atlas and anatomical boundaries). We provide the methods as the MagellanMapper software and the eight 3D reconstructed ADMBA atlases. These resources facilitate whole-organ quantitative analysis between samples and across development.
Article activity feed
-
###Reviewer #2:
Despite the availability of a high resolution, expertly annotated digital adult mouse brain atlas (Allen CCFv3), accurately labeled 3D digital atlases across mouse neural development are lacking. The authors have filled that gap by developing novel computation methods that transform slice annotations in the Allen Developing Mouse Brain Atlas into digital 3D reference atlases. They demonstrate that the resulting brain parcellations are superior to a naive agglomeration of the existing 2D labels, and provide MagellanMapper, a suite of tools to aid quantitative measures of brain structure. Cellular level whole-brain quantitative analysis is rapidly becoming a reality in many species and this manuscript provides a foundational resource for mouse developmental studies. The methods are sophisticated, carefully applied and …
###Reviewer #2:
Despite the availability of a high resolution, expertly annotated digital adult mouse brain atlas (Allen CCFv3), accurately labeled 3D digital atlases across mouse neural development are lacking. The authors have filled that gap by developing novel computation methods that transform slice annotations in the Allen Developing Mouse Brain Atlas into digital 3D reference atlases. They demonstrate that the resulting brain parcellations are superior to a naive agglomeration of the existing 2D labels, and provide MagellanMapper, a suite of tools to aid quantitative measures of brain structure. Cellular level whole-brain quantitative analysis is rapidly becoming a reality in many species and this manuscript provides a foundational resource for mouse developmental studies. The methods are sophisticated, carefully applied and thoroughly evaluated. I have mostly minor comments that should be interpreted as suggestions to strengthen or clarify the presentation, not an indication of any significant concerns.
The authors developed a clever 'edge-aware procedure' that they first employed to extend existing labels to unannotated lateral regions of the brain, taking advantage of intensity gradations in underlying microscope images. As this is an innovative procedure, the authors should manually annotate a small part of the lateral brain region to compare accuracy and compare computationally generated labels to the partial lateral labels in P28 brain.
I have questions about how well the edge-aware procedure performed internally within the brain to smooth region parcellation. First, the edge-aware procedure relies on intensity differences in the light microscope images. However, the work of neuroanatomists would be dramatically simplified if such gradations provided sufficient information for brain segmentation. Annotations present in the ADMBA took advantage of co-aligned ISH data (and computational approaches using co-aligned gene expression data have been used for de novo brain parcellation). Intensity differences in the light-microscope images may not always provide enough information for accurate segmentation. Could there be instances where adjacent regions do not have intensity differences, and the edge-aware procedure actually reduces the accuracy of the manual annotation? Second, it does appear that despite the care to avoid losing thin structures, there is some loss, for example for the light-green structure in the forebrain in Fig. 5E. Could the authors indicate if all labels were preserved, and perhaps provide information on volume changes by label size.
The accuracy of non-rigid registration of light-sheet images to the references is assessed only using a DSC value for whole-brain overlaps. This does not assess the precision of registration within the brain. The authors should apply some other measure to measure the quality of alignment within the brain (e.g. mark internal landmarks visible in the reference and original light-sheet images, and measure the post-registration distance between them).
The P56 reference is close to an adult brain. The authors should compare the boundaries of their computationally derived parcellations to the recently published Allen CCFv3 brain regions.
-
###Reviewer #1:
The manuscript demonstrated some interesting aspects of the data processing for the 3D registration of the mouse brain. At the same time, several concerns need to be addressed, by either revising the text or making additional computations.
The 3D "smoothing" was the central part of the method reported in the manuscript. For example, the inclusion of the "skeletonization" step helped prevent the loss of thin structures compared to the previous methods such as the one by Niedworok et al (Ref #40 in the manuscript). However, the overall improvement did not involve any conceptually new algorithm but instead relied on the optimization of known parameters, which may appear incremental. The authors should avoid overstating their work.
The pipeline of the method involved the "mirroring" before the "smoothing" steps. Is it …
###Reviewer #1:
The manuscript demonstrated some interesting aspects of the data processing for the 3D registration of the mouse brain. At the same time, several concerns need to be addressed, by either revising the text or making additional computations.
The 3D "smoothing" was the central part of the method reported in the manuscript. For example, the inclusion of the "skeletonization" step helped prevent the loss of thin structures compared to the previous methods such as the one by Niedworok et al (Ref #40 in the manuscript). However, the overall improvement did not involve any conceptually new algorithm but instead relied on the optimization of known parameters, which may appear incremental. The authors should avoid overstating their work.
The pipeline of the method involved the "mirroring" before the "smoothing" steps. Is it possible to perform the "smoothing" of one hemisphere and then "mirror" the smoothed 3D atlas onto the other hemisphere to check for the alignment? By doing so, the other hemisphere could serve as an internal control for the quality and accuracy of the 3D atlas.
3.) The "edge-aware" adjustment, which was essential for the improvement of 3D atlas, surely worked for the large brain regions with identifiable anatomical edges based on the 2D images. However, for more delicate subregions (e.g., those in the hypothalamus) without clear anatomical boundaries, this adjustment step may become ineffective. What could then be done for these subregions? Also, it is important to note that the anatomical edges required the manual annotation.
The results presented throughout the manuscript are the axial views of brains. It would be informative to include, at least in Figures 2 and 3, the coronal views of 3D atlases to exemplify the quality.
It is unclear why the authors chose the P0 brains for the lightsheet imaging. In addition, since both male and female mice were analyzed, is there any difference observed within the 3D brain atlases obtained?
-
##Preprint Review
This preprint was reviewed using eLife’s Preprint Review service, which provides public peer reviews of manuscripts posted on bioRxiv for the benefit of the authors, readers, potential readers, and others interested in our assessment of the work. This review applies only to version 3 of the manuscript. Joseph G Gleeson (Howard Hughes Medical Institute, The Rockefeller University) served as the Reviewing Editor.
###Summary:
Despite the availability of a high resolution, expertly annotated digital adult mouse brain atlas (Allen CCFv3), accurately labeled 3D digital atlases across mouse neural development are lacking. The authors have filled that gap by developing novel computational methods that transform slice annotations in the Allen Developing Mouse Brain Atlas into digital 3D reference atlases. They demonstrate that …
##Preprint Review
This preprint was reviewed using eLife’s Preprint Review service, which provides public peer reviews of manuscripts posted on bioRxiv for the benefit of the authors, readers, potential readers, and others interested in our assessment of the work. This review applies only to version 3 of the manuscript. Joseph G Gleeson (Howard Hughes Medical Institute, The Rockefeller University) served as the Reviewing Editor.
###Summary:
Despite the availability of a high resolution, expertly annotated digital adult mouse brain atlas (Allen CCFv3), accurately labeled 3D digital atlases across mouse neural development are lacking. The authors have filled that gap by developing novel computational methods that transform slice annotations in the Allen Developing Mouse Brain Atlas into digital 3D reference atlases. They demonstrate that the resulting brain parcellations are superior to a naive agglomeration of the existing 2D labels, and provide MagellanMapper, a suite of tools to aid quantitative measures of brain structure. Cellular level whole-brain quantitative analysis is rapidly becoming a reality in many species and this manuscript provides a foundational resource for mouse developmental studies. The methods are sophisticated, carefully applied and thoroughly evaluated. The manuscript reports a computational approach to transforming available 2D atlases of mouse brains into the 3D volumetric datasets. By optimizing the "smoothing" steps, a better quality of such 3D atlases is produced. In addition, the authors applied their method to the imaging dataset of neonatal mouse brains obtained by lightsheet microscopy, as proof of its potential utilization in research.
-
