Reconstruction of cell lineages and behaviors underlying arthropod limb outgrowth with multi-view light-sheet imaging and tracking

During development coordinated cell behaviors orchestrate tissue and organ morphogenesis to suit the lifestyle of the organism. We have used here the crustacean Parhyale hawaiensis to study the cellular basis of limb development. Transgenic Parhyale embryos with fluorescently labeled nuclei were imaged at high spatiotemporal resolution with multi-view light-sheet fluorescence microscopy over several days of embryogenesis spanning appendage morphogenesis from early specification up to late differentiation stages. Cell tracking with a new tool called Massive Multi-view Tracker (MaMuT) enabled the reconstruction of the complete cell lineage of an outgrowing thoracic limb with single-cell resolution. In silico clonal analyses suggested that the limb primordium becomes subdivided from an early stage first into anterior-posterior and then into dorsal-ventral compartments whose boundaries intersect at the distal tip of the growing limb. Limb bud formation is associated with the spatial modulation of cell proliferation, while limb elongation is also driven by the preferential orientation of division of epidermal cells along the proximal-distal axis of growth. Cellular reconstructions were predictive of the expression patterns of limb development genes including the Decapentaplegic (Dpp) morphogen. HIGHLIGHTS Multi-view light-sheet microscopy of crustacean embryos from species Parhyale hawaiensis are ideal for cellular-level analysis of organ morphogenesis. Lineages of 3-dimensional organs were reconstructed at single-cell resolution with the Fiji/ImageJ plugin Massive Multi-view Tracker. The Parhyale limb primordium undergoes early lineage restrictions associated with particular cell behaviors and patterns of gene expression. Differential rates of cell proliferation and oriented cell divisions guide appendage proximal-distal outgrowth.


INTRODUCTION
Developmental morphogenesis is the formation of tissues and organs with particular sizes and shapes to suit the lifestyle of multicellular organisms (Lecuit and Le Goff, 2007).
Morphogenesis is driven by patterned cell activities. Therefore, only detailed descriptions of cell lineages and cell behaviors can provide a firm ground for any morphogenetic analysis (Buckingham and Meilhac, 2011). Furthermore, morphogenesis requires capturing and integrating information across multiple levels of biological organization: from the subcellular level, to cell behaviors, to emerging biological form (Keller, 2013;Liu and Keller, 2016).
In this study, we have focused on the crustacean amphipod Parhyale hawaiensis that satisfies a number of appealing biological and technical requirements as an experimental model system to study appendage (limb) morphogenesis at single-cell resolution from early specification until late differentiation stages (Stamataki and Pavlopoulos, 2016).
First, Parhyale is a direct developer; most aspects of the adult body plan, including appendages, are specified during the 10 days of embryogenesis when imaging is readily possible (Browne et al., 2005). Second, Parhyale exhibits a striking morphological gradation along the main body axis; each embryo develops a variety of specialized appendages along the anterior-posterior axis (e.g. antennae, mouthparts, limbs for defense, walking, swimming etc.) that differ in size, shape and pattern (Martin et al., 2016;Pavlopoulos et al., 2009;Wolff and Scholtz, 2008). Third, Parhyale eggs have the appropriate size and optical properties for microscopic live imaging of constituent cells with very high spatial and temporal resolution; embryos are about 500 µm long, transparent with low autofluorescence and light scattering. Finally, an increasing number of functional genetic approaches, embryological treatments, genomic and transcriptomic resources allow diverse experimental manipulations of Parhyale embryos (Kao et al., 2016). Transient fluorescent labeling by injecting early stage Parhyale embryos with mRNAs encoding fluorescent markers coupled with standard epifluorescence or confocal microscopy has been widely adopted to study early processes like gastrulation and germband formation (Hannibal et al., 2012). However, to achieve a comprehensive coverage of appendage formation, Parhyale embryos need to be imaged ideally from multiple angular viewpoints and continuously from day 3 up to at least day 8 of embryogenesis (Browne et al., 2005). In this article, we demonstrate that transgenic Parhyale embryos with fluorescently labeled nuclei can be imaged routinely for several consecutive days using Light-sheet Fluorescence Microscopy (LSFM). LSFM is an ideal technology for studying how cells form tissues and organs in intact developing embryos (Huisken et al., 2004;Keller et al., 2008;Truong et al., 2011). It enables biologists to capture fast and dynamic developmental processes at very high spatiotemporal resolution, over long periods of time, and with minimal bleaching and photo-damage (Khairy and Keller, 2011;Schmied et al., 2014;Weber et al., 2014). In addition, samples can be optically sectioned from multiple angles (multi-view LSFM) that can be combined computationally to reconstruct the entire specimen with a more isotropic resolution (Chhetri et al., 2015;Krzic et al., 2012;Swoger et al., 2007;Tomer et al., 2012;Wu et al., 2013).
Although the amount and type of data generated by multi-view LSFM raise several challenges for image processing and analysis, most of them have been efficiently addressed. Software solutions exist for registration of acquired views (z-stacks) in each time-point (Preibisch et al., 2010), for fusion of views into a single output z-stack with nearly isotropic resolution (Preibisch et al., 2014), and for rendering and three dimensional reconstruction of the entire imaged volume (Pietzsch et al., 2015). These processes can be repeated automatically using high-performance computing resources for several hundred or thousands of time-points to generate a four-dimensional representation (three spatial dimensions + temporal dimension) of the embryo as it develops over time Schmied et al., 2014). Automated approaches for cell segmentation and tracking have also been developed , however they do not yet reach the precision required for unsupervised extraction of cell lineages. To address this issue, we describe here the Massive Multi-view Tracker (MaMuT) software that allows the visualization, annotation, and accurate lineage reconstruction of large multi-dimensional microscopy data.
We quantitatively followed cell dynamics with the MaMuT software on the Parhyale multi-view LSFM acquisitions in order to understand the cellular basis of arthropod appendage morphogenesis. As revealed by lineage tracing experiments in the leading arthropod model Drosophila melanogaster, the leg and wing primordia become progressively subdivided into distinct cell populations (called compartments when lineage-restricted), first along the anterior-posterior (AP) axis during their specification in the early embryo, and later along the dorsal-ventral (DV) axis during larval stages (Dahmann et al., 2011;Garcia-Bellido et al., 1973;Steiner, 1976). These lineagedependent or independent subdivisions of the tissue acquire distinct cell fates or identities driven by domain-specific expression of patterning genes (called selector genes if lineally inherited), as well as by the localized expression of signaling molecules at compartment boundaries ("organizers") that control patterning and growth of the developing organs (Garcia-Bellido, 1975;Lawrence and Struhl, 1996;Mann and Carroll, 2002;Restrepo et al., 2014). Besides these regionalization mechanisms, oriented cell divisions have been implicated as a general mechanism in shaping the Drosophila wing and other growing organs in diverse developmental systems (Baena-Lopez et al., 2005). Other cellular activities like patterned cell proliferation and cell rearrangement could also play a role in the formation of limb buds and their elongation along the proximal-distal axis, but these processes have been difficult to study in existing arthropod models. By acquiring the lineage information at single-cell resolution, as well as the positional coordinates and their change over time for all constituent cells in a developing limb, we identified the lineage restrictions and morphogenetic cellular behaviors operating during Parhyale limb bud formation and elongation, and compared these to the Drosophila and other arthropod paradigms. Finally, we validated our reconstructions by studying at cellular resolution the expression of Decapentaplegic (Dpp) signaling components implicated in limb patterning and growth.

Imaging Parhyale embryogenesis with multi-view LSFM
6 Segment formation and maturation in Parhyale occurs sequentially in anterior-toposterior progression (Browne et al., 2005). In our LSFM recordings, we were particularly interested in capturing appendage development in the anterior thorax of Parhyale embryos. In these segments, appendages were specified at about 3.5 days after egg-lay (AEL) at 25˚C. Over the next 4 days, appendage buds bulged out ventrally, elongated along their proximal-distal axis and became progressively segmented until they acquired their definite morphology at around 8 days AEL (Figure 1).
We generated transgenic Parhyale embryos with fluorescently labeled nuclei as described in Materials & Methods. Individual 3-day old embryos (stage S13) were mounted for LSFM in a column of low melting agarose with scattered red fluorescent beads. Embryos were imaged on a Zeiss Lightsheet Z.1 microscope that offers dual-sided illumination with single-sided detection. Several parameters (detailed in Materials & Methods) were optimized to cover all stages of Parhyale appendage development over 4-5 days of embryogenesis at single-cell resolution with adequate temporal sampling and with minimal photo-bleaching and damaging. A typical 4 to 5-day long Parhyale embryo recording was composed of more than 1 million images resulting in >7 TB datasets.
The slow tempo of Parhyale development enabled LSFM imaging of the entire embryo from multiple highly overlapping views (every 45˚, Figure 1A) without noticeable displacement of cell nuclei between all the views acquired in each time-point. Development of the entire embryo was reconstructed from the raw input views according to the following steps (detailed in Materials & Methods) using open-source software available as plug-ins in the Fiji (Fiji Is Just ImageJ) biological image analysis platform (Schindelin et al., 2012): 1) image file preprocessing, 2) bead-based spatial registration of views within one time-point, 3) fusion by multi-view deconvolution, 4) bead-based temporal registration across time-points, 5) computation of temporally registered fused volumes, and 6) 4D rendering of the spatiotemporally registered fused data (Preibisch et al., 2014;Preibisch et al., 2010;Schmied et al., 2014). This processing resulted in almost isotropic resolution of fused volumes and was used for visualization of Parhyale embryo development with cellular resolution ( Figure 1B).

Following appendage morphogenesis in 4D reconstructed Parhyale embryos
Appendage morphogenesis in arthropods involves patterning, growth and differentiation of ectodermal cells organized in an epithelial monolayer that will give rise to the appendage epidermis and will secrete the cuticle. Our study focused on the dynamics of the Parhyale ectoderm from the stages of appendage specification (day 3 AEL) up to differentiation (day 8 AEL). During germband formation, the ectoderm contributing to the posterior head and the trunk (known as post-naupliar germband) became organized in a stereotyped grid-like pattern with a very precise arrangement of rows and columns of cells ( Figures 1C-1D (Browne et al., 2005;Dohle et al., 2004).
Each row of cells corresponded to one parasegment, which is the unit of early metameric organization in Parhyale embryos, like in Drosophila and other arthropods Scholtz et al., 1994). Two rounds of longitudinally-oriented cell divisions differentiating along their proximal-distal axis ( Figures 1G-1K). The gnathal appendages (mandible, maxilla 1, maxilla 2) grew relatively little in size, remained unsegmented and bifurcated heralding the development of gnathal palps ( Figure 1H). The most conspicuous appendages of the embryo were the two antennae and the thoracic appendages. These appendages elongated considerably along their proximal-distal axis and became progressively subdivided into distinct segments (Figures 1F-1K and Figure   6). At the end of the imaging period, morphogenesis appeared nearly complete. All thoracic appendages (T1 to T8) were composed of 7 constituent segments that had distinct pattern, size and shape to serve their specialized function: T1 maxillipeds for 8 feeding, T2 and T3 gnathopods for grasping and T4 to T8 legs for locomotion. Besides elongation and segmentation, thoracic appendages also developed a number of different proximal outgrowths. The maxillipeds developed two ventral outgrowths, called endites ( Figure 1H), while the more posterior thoracic appendages developed dorsal outgrowths, called epipodites, serving as protective coxal plates and respiratory gills ( Figures 1I-1K).
Thus, multi-view LSFM imaging captures the entire gamut of differential appendage morphogenetic events along the body axis of the Parhyale embryo in a single time-lapse experiment.

MaMuT: a platform for cell tracking in multi-view and multi-terabyte datasets
In order to examine the cellular basis of morphogenesis, we developed a novel Fiji software application to extract cell lineages from multi-view and multi-terabyte datasets.
This tool was dubbed MaMuT for Massive Multi-view Tracker and is a hybrid and extension of two existing Fiji plugins: the BigDataViewer visualization engine (Pietzsch et al., 2015); http://imagej.net/BigDataViewer) and the TrackMate annotation engine (Tinevez et al., 2016); http://imagej.net/TrackMate). MaMuT is an interactive, userfriendly tool for visualization, annotation, object tracking and lineage reconstruction of large multi-dimensional microscopy data. Within the Fiji ecosystem, MaMuT is tightly integrated with the plugins for multi-view LSFM data processing ( Figure 3A). Documentation, training datasets and tutorials on MaMuT can be found at the dedicated link http://imagej.net/MaMuT.
MaMuT can handle multiple data sources but was developed primarily to enable the analysis of large and complex LSFM image datasets (Figure 3 and Figure S1). Its unique feature is the ability to annotate an image volume synergistically from all available input views without the need to fuse these into a single volume. All raw views were first registered and then imported into MaMuT ( Figure 3A). The software allowed users to open the raw image data in as many Viewer windows as required and visualize each zstack in any desired orientation, scale and time-point ( Figures 3B-3D). Most importantly, all Viewer windows were synced based on the calculated registration parameters and shared a common physical coordinate system. That is, upon selecting an object of interest, like a cell or nucleus ("spot") in one Viewer, the same spot was identified and displayed in all other windows, and its x, y, z position was mapped onto this common physical space ( Figures 3B-3D). This unique functionality of MaMuT allowed us to identify and track all constituent cells in a developing appendage continuously from the early germband stages until the later stages of 3D organ outgrowth (Figures 4 to 6), when the information from multiple views was required to fully reconstruct the appendage with single-cell resolution. Selected nuclei were tracked over time in user-defined time intervals, and the reconstructed trajectories and lineages were also displayed in two additional synced windows, the TrackScheme and 3D Viewer. The TrackScheme lineage browser and editor displayed the reconstructed cell lineage tree with tracked nuclei Although, all lineage reconstructions presented in this article were generated manually, the latest MaMuT architecture also offers two functionalities for automated tracking: i) a semi-automated option where individual nuclei can be selected by the user and segmented and tracked computationally over time, and ii) the option to import into MaMuT fully automated annotations generated by the Tracking with Gaussian Mixture Models (TGMM) software , which is one of the most accurate and computationally efficient methods for segmentation and tracking of fluorescently labeled nuclei ( Figure 3A). Therefore, MaMuT is a versatile platform that can be used either for fully manual or semi-automated tracking of selected populations of objects of interest, or for visualization and editing of fully automated computational predictions for systemslevel lineage reconstructions.

Single-cell lineage reconstruction of a Parhyale thoracic limb
We next deployed the manual version of MaMuT to extract the developmental lineage tree of a Parhyale thoracic limb. By convention, ectodermal rows in the Parhyale germband are identified from anterior to posterior by ascending numbers E0, E1, E2 etc. In accordance with previous studies in malacostracan crustaceans and various other arthropods, our reconstructions demonstrated that each Parhyale thoracic limb consisted of cells from two neighboring parasegments (Browne et al., 2005;Dohle et al., 2004;Scholtz et al., 1994;Wolff and Scholtz, 2008). The T2 limb that we analyzed in-depth ( Figure 4A-4E) developed from rows b, c and d of the E4 parasegment and from rows a and b of the following E5 parasegment ( Figure 4F and Figure S2). Cells that arose from rows c, d and a occupied the entire length of the limb proximal-distal axis and the body wall part of the T2 segment ( Figure S2). Descendent cells from rows b contributed only to the proximal limb and intersegmental territories ( Figure S2). Cells in medial columns 1 and 2 gave rise to the nervous system and sternites and were not considered further in this study. The more lateral columns 3 to 9 contributed the epidermal cells forming the limb ( Figure S3).
We fully tracked 34 cells constituting the T2 limb primordium over 50 hours of Parhyale embryogenesis giving rise to a total of 361 epidermal cells. We started tracking each of these 34 cells as they divided longitudinally from the 2-row to the 4-row parasegment and monitored them continuously during the subsequent rounds of divisions (referred to as differential divisions, DDs). The number of DDs observed during these 50 hours (400 time-points) varied dramatically between tracked cells from just one DD in the slowest dividing lateral cells of the T2 primordium (cells E4b8, E5a9/b9) to five DDs in the fastest dividing central cells (cells E4c3-c6 and E4d3-d6). Although the clonal composition of crustacean appendages had been described previously using single-cell injections with lypophilic dyes (Wolff and Scholtz, 2008), the reconstruction presented here is the most comprehensive lineage tree for any developing arthropod limb published to date ( Figure S4).

Early lineage restrictions along the anterior-posterior and dorsal-ventral axes
We first asked whether these complete reconstructions could reveal any lineage-based subdivisions in the developing limb. The AP restriction at the border of neighboring parasegments from the 1-row-parasegment stage onwards has been revealed in Parhyale and other malacostracan embryos by classic embryological descriptions, lineage tracing experiments and expression studies for the segment polarity gene engrailed that marks the posterior compartment (Browne et al., 2005;Dohle et al., 2004;Hejnol and Scholtz, 2004;Scholtz et al., 1994). In agreement with this AP restriction, during limb specification, outgrowth and elongation there was a straight clonal boundary running between the anterior compartment cells derived from the E4b, c and d rows and the posterior compartment cells derived from the E5a and b rows ( Figure 4 and Figure S2).
Having established that the well-known AP boundary was readily observable in our complete lineages, we next sought to identify any subdivision along the dorsal-ventral axis. Compartments were classically discovered by clonal analysis using mitotic recombination. In our reconstructions, we could generate clones digitally from arbitrary cells at different stages of appendage development. We reasoned that we could reveal the timing and position of any heritable DV restriction by piecing together correctly all founder cells of dorsal or ventral identity in a way that the two polyclones (i.e. compartments) would stay separate and form a lasting straight interface between them. This analysis suggested that there is indeed a DV separation that took place at the 4-rowparasegment stage; the DV boundary ran between the E4b and c rows anteriorly, between the E5a and b rows posteriorly, and between cells E4c4-c5, E4d3-d4 and E5a4-a5 medially ( Figure 4F). Throughout limb development, the dorsal and ventral cells formed a sharp boundary between themselves extending along the proximal-distal axis (Figures 4I-4J' and Figure S2). In order to evaluate the reproducibility and stereotypy of the AP and, in particular, the DV separation across Parhyale embryos, we analyzed an independently imaged and reconstructed T2 limb from a different embryo ( Figure S5).
We confirmed that four identical compartments (anterior-dorsal, anterior-ventral, posterior-dorsal and posterior-ventral) could be derived in this independent reconstruction with straight boundaries and no cell mixing between neighboring compartments (compare Figure 4 with Figure S5).
Taken together, these results suggested that in silico analyses of comprehensive and accurate lineage reconstructions can provide novel insights into clonal subdivisions and the underlying developmental patterning mechanisms in species where sophisticated genetic dissections are not implemented yet. reconstructions suggested that cell movements played little role, at least during the analyzed stages of limb outgrowth. We identified few instances of cells intercalating between one another along the proximal-distal axis during the extension phase slightly contributing to the narrowing and lengthening of the growing limb ( Figure S6).

Quantification of differential cell behaviors during limb formation
We next asked whether our reconstructions could provide insights into the morphodynamics of limb formation at a finer scale. Two cell behaviors implicated in tissue and organ morphogenesis were readily quantifiable in our nuclear trackings, namely the pattern of cell proliferation and the orientation of cell divisions (Baena-Lopez et al., 2005;Boehm et al., 2010). These cell activities have been traditionally inferred in developing tissues from the distribution, size and shape of somatic clones induced by various means (Buckingham and Meilhac, 2011). This approach could be also adapted here by generating in silico clones for every single tracked epidermal cell between any developmental stages of interest ( Figure S7). Yet, the MaMuT reconstructions enabled us to enrich the lineage information with rigorous quantitative analyses of the rate and orientation of mitotic divisions using the extracted x, y, z, t coordinates for all tracked nuclei.  Figures 4M-4N). Another stripe of faster dividing cells was discovered in the anterior-dorsal cells abutting the AP compartment boundary ( Figures 4N-4N').
Next, we looked for any biases in the orientation of mitotic divisions that could be associated with Parhyale limb morphogenesis similar to the Drosophila wing paradigm (Baena-Lopez et al., 2005). The AP compartment boundary ran parallel to the proximaldistal axis of limb growth and provided an excellent axis of reference against which we compared the axis of each cell division (Figures 5). As expected, all early divisions in the T2 primordium were perpendicular to the AP compartment boundary (i.e. they were parallel to the AP axis) confirming the strict longitudinal orientation of row divisions ( Figure 5F). Cell divisions acquired a more heterogeneous pattern after the 4-rowparasegment stage ( Figure 5G). An increasing number of mitotic spindles aligned progressively along the proximal-distal axis during limb bud formation ( Figure 5H) and elongation ( Figures 5I-5J).
Collectively, the information extracted from our spatiotemporally resolved lineage trees strongly suggested that Parhyale limb outgrowth is driven by at least two patterned cell behaviors: the differential rates of cell proliferation and the orderly arrangement of mitotic spindles.

Cellular basis of the elaboration of the limb proximal-distal axis
We then asked whether our dataset could help resolve the patterning mechanisms operating during appendage segmentation. Appendage segmentation is a gradual process as the elongating proximal-distal axis becomes progressively subdivided into an increasing number of elements ( Figure 6) (Rauskolb, 2001). At the tissue level, the This lineage analysis demonstrated that the cells that gave rise to the proximal, medial and distal limb segments occupied distinct mediolateral positions in the germband grid at the 4-row-parasegment stage ( Figure 6M) and distinct proximal-distal positions in the early limb bud ( Figure 6N). Later on, analysis of the bipartite limb indicated that the proximal element gave rise to the proximal segments coxa, basis and ischium, while the distal element gave rise to the distal segments merus, carpus, propodus, and dactylus ( Figure 6O). In other words, the first segmental subdivision occurred between ischium/merus and was followed by the basis/ischium subdivision ( Figure 6P), the propodus/dactylus, carpus/propodus and coxa/basis subdivisions ( Figure 6Q), and the carpus/merus subdivision ( Figure 6R). Tracking analysis indicated that the cells forming the distal segments merus, carpus, propodus, and dactylus were not related by lineage, but by position as they originated from a medial territory around the intersection of the 4 compartments ( Figure S9). During the subsequent limb elongation stages, these distal cells kept separate from more proximal cells at the prospective ischium/merus joint, suggesting that limb segments may represent secondary tissue subdivisions along the proximal-distal axis ( Figure S9) (Milan and Cohen, 2000).

Expression of limb patterning genes validates cellular models of Parhyale limb morphogenesis
To test the validity of our cellular models and make a first link between expression of limb patterning genes and morphogenetic cell behaviors, we cloned and analyzed by in situ hybridization the expression of the Parhyale decapentaplegic (Ph-dpp) gene that encodes a member of the Bone Morphogenetic Protein 2/4 class of signaling molecules.
In Drosophila imaginal discs, Dpp signaling controls the dorsal cell fate in the leg, as well as growth via cell proliferation in the wing (Akiyama and Gibson, 2015;Brook and Cohen, 1996;Harmansa et al., 2015;Rogulja and Irvine, 2005;Svendsen et al., 2015).
Therefore, probing Ph-dpp expression in forming Parhyale limb buds could provide a direct test for our cell-based predictions regarding the DV lineage restriction and the differential cell proliferation rates in the limb primordium.
Analysis of S18 embryos (96 hrs  To get an insight into the downstream effects of Dpp signaling in the Parhyale limb buds, we also analyzed by in situ hybridization the expression of the Tbx6/Dorsocross (Doc) gene that responds to high levels of Dpp signaling in the dorsal region of the Drosophila embryo and leg disc (Reim et al., 2003;Svendsen et al., 2015). Expression of the single Collectively, these results demonstrate how the reconstruction of cell lineages and behaviors can provide solid predictions and powerful contexts to study the expression and function of associated genes.

DISCUSSION
In this article, we have established an integrated framework to study the cellular and genetic basis of developmental morphogenesis. By combining light-sheet microscopy with a newly developed versatile software for cell tracking and lineage reconstruction in large multi-view datasets, we have revealed the cellular architecture and cellular dynamics underlying organ morphogenesis in a non-conventional experimental model, and have provided new insights into the mechanisms underlying the three-dimensional epithelial remodeling during arthropod limb morphogenesis.

MaMuT, a user-friendly tool for lineaging multi-dimensional image volumes
MaMuT's unique feature that is currently not available in any other commercial or opensource software platforms for image data visualization and analysis is the capacity to track objects synergistically from all available views in multi-view LSFM recordings.
This functionality has a number of very important advantages. Raw image stacks do not have to be fused into a single volume, which is computationally by far the most demanding step in multi-view LSFM image processing requiring access to high-end workstations and computer clusters or down-sampling of image volumes (Preibisch et al., 2014). The users also preserve the original redundancy of their data, which in many cases, including Parhyale, allows capturing each object of interest from two or more neighboring views that can be interpreted independently for a more accurate analysis.
Finally, MaMuT allows users to analyze sub-optimal datasets that cannot be fused properly or may create fusion artifacts. Of course, combining the raw views with a highquality fused volume is the best available option, especially when handling complex datasets with high cell densities.
While offering multi-view tracking, MaMuT delivers also a number of other important functionalities for the exploration of multi-dimensional datasets. First, MaMuT is a userfriendly turnkey software with a convenient interface for interactively navigating, inspecting and curating the image and annotation data. Any time-lapse series acquired by any microscopy modality that can be opened in Fiji can be also imported into MaMuT.
Second, MaMuT offers a highly responsive and interactive navigation through multiterabyte datasets fostered by the BigDataViewer architecture for image storage, loading and caching (Pietzsch et al., 2015). Individual z-stacks representing different views and/or channels of a multi-dimensional image dataset can be displayed independently (or in any combination) in multiple Viewer windows. Each Viewer can be also adjusted independently for color and brightness, zoom, translation and rotation of image stacks and browsing through time points. Third, MaMuT relies on TrackMate for the annotation model, which is a Fiji plugin developed for tracking purposes in conventional imaging datasets (Tinevez et al., 2016). Objects of interest like cells and nuclei ("spots") can be selected synergistically from all available Viewers and followed over multiple time points to reconstruct their trajectories ("tracks"), division patterns and lineage information. Fourth, the created spots and tracks can be visualized and edited interactively in the Viewers and the TrackScheme lineage browser, and can be animated in the 3D Viewer window. For visual interpretation of the data, annotations can be colored based on the primary x, y, z, t and lineage information or derived numerical parameters such as velocity, displacement, division time and others. Fifth, lineages can be reconstructed in a fully manual, semi-automated or fully automated manner followed by manual curation if necessary. Thus, MaMuT can address various experimental designs ranging from analyzing a small subset of objects in the imaged volume to systems-wide analyses of all constituent parts. Sixth, all spot and track information is stored in an xml file that can be exported from MaMuT to other interfaces for more specialized analyses. Seventh, decentralized annotation by multiple users has been made possible by also developing a web service for remote access to large image volumes stored online. Last but not least, following on the tradition of the Fiji community for open-source distribution of biological image analysis software, MaMuT is provided freely and openly to the community, it is extensively documented and can be customized by other users.

Multi-view LSFM is ideal to reveal the cellular basis of Parhyale morphogenesis
The LSFM imaging technology is empowering biologists to study developmental processes with unprecedented spatial and temporal resolution (Keller, 2013;Khairy and Keller, 2011;Schmied et al., 2014;Weber et al., 2014). We demonstrated here that in the multi-view mode it is particularly suited for imaging embryos of the crustacean amphipod The tempo and mode of development has also important ramifications for Parhyale multi-view LSFM imaging and tracking. First, the slow tempo of development enabled to image the embryo at a very high spatial resolution through the acquisition of multiple and highly overlapping views without compromising the temporal resolution. Thus, Parhyale recordings can match or exceed the spatiotemporal resolution of other pioneering models in modern LSFM microscopy, like Drosophila and zebrafish, even when access to highest-speed instruments is not available. Second, due to the optical clarity of the embryo and positioning of the appendages on the surface of the developing embryo, all constituent cells were captured and tracked for systems-level quantitative analyses.
Finally, the conserved, stereotypic and highly ordered organization of the post-naupliar ectoderm in Parhyale will allow to identify homologous cells and compare lineages, cell behaviors and associated genes between serially homologous structures in the same embryo, across embryos and even across malacostracan crustaceans.

Cellular basis of arthropod limb morphogenesis: lessons from Parhyale
All our knowledge on the cellular basis of arthropod limb morphogenesis comes from studies in Drosophila (Fristrom, 1988;von Kalm et al., 1995). However, development of Drosophila limbs is extremely derived and not representative for many insects, much less arthropods in general. Drosophila limb specification, patterning, growth, and differentiation take place at distinct developmental stages during embryonic, larval and pupal development. On the contrary, all these processes come about during embryogenesis in most other arthropods, including Parhyale, and limbs are direct threedimensional outgrowths of the embryonic body wall.
Using lineage-dependent or independent mechanisms, initially homogeneous fields of cells become progressively subdivided into distinct domains that differ in gene expression and often form and maintain sharp boundaries between themselves. Lineagebased subdivisions have been identified in the Drosophila wing and leg primordia during their compartmentalization along the AP axis in early embryogenesis, and during DV compartmentalization of the wing disc during early larval development (Garcia-Bellido et al., 1973;Steiner, 1976). These heritable subdivisions are based on the compartmentspecific expression of selector genes and localized induction of signaling molecules ("organizers") at the AP compartment boundaries in both legs and wings and the DV boundary of the wing (Dahmann et al., 2011;Garcia-Bellido, 1975;Lawrence and Struhl, 1996;Mann and Carroll, 2002). The mechanism underlying DV separation in the Drosophila leg disc is not clear yet. Classical lineaging experiments suggested a heritable DV division during larval development, although the clonal restriction was not entirely strict and was only detected in the anterior compartment but not the posterior one (supposedly due to its smaller size) (Steiner, 1976). It is also not clear yet whether the domains of expression of dorsal and ventral selector genes in the leg disc are coextensive with their corresponding compartments, and what are the cellular mechanisms contributing to regionalization of the leg along its DV axis (Brook and Cohen, 1996;Svendsen et al., 2015).
Beyond Drosophila, our understanding of the AP and DV organization in other arthropod limbs has relied so far entirely on gene expression studies. Expression of segment polarity genes, like engrailed and wingless, has demonstrated that the AP separation is widely conserved across the arthropods and takes place during the segmentation stages (Angelini and Kaufman, 2005;Patel et al., 1989). In Parhyale, the AP compartment boundary is established at the 1-row stage at the interface of neighboring parasegments (Browne et al., 2005;Scholtz et al., 1994). With the exception of descriptive gene expression studies (Janssen et al., 2008;Prpic et al., 2003), the mechanism, timing and position of the DV separation in arthropod limbs has remained unexplored at the cellular level. This limitation stems from the lack of methodologies to systematically generate and interpret the distribution and shape of mitotic recombination clones in non-model arthropods. Despite this limitation, we have been able to explore the cellular mechanisms underlying the developmental subdivisions in the Parhyale limb by analyzing the dynamics of digital clones in reconstructed T2 limbs. First, we confirmed that this approach was indeed successful in revealing the position and timing of the AP compartment boundary; the population of anterior cells (descendants from rows E4b, c and d) remained together and separate by a straight boundary from posterior cells (descendants from rows E5a and b). Second, by applying the same rules we revealed a putative heritable subdivision along the DV axis in two independently imaged and analyzed T2 limbs. The presumptive DV compartment boundary formed at the 4-rowparasegment stage between the E4b and c rows anteriorly, the E5a and b rows posteriorly, and between cells E4c4-c5, E4d3-d4 and E5a4-a5 medially. Third, expression of Distal-less (Dll), which is one of the earliest markers of limb specification, initiates at the 4row-parasegment in the cells contributing to the appendage primordia (Browne et al., 2005;Hejnol and Scholtz, 2004). Dll expression is first detected in the d4 cell, which was one of the cells at the intersection of the AP and DV compartment boundaries. Fourth, this intersection coincided with the leading tip of the forming limb throughout epithelial remodeling and outgrowth.
These results demonstrate that the Parhyale limb perfectly conforms to the boundary model originally proposed by Hans Meihardt to whose memory this article is dedicated (Meinhardt, 1983). Meinhardt's model postulated that a secondary developmental field, i.e. the proximal-distal axis of a limb that is specified during embryogenesis de novo relative to the main AP and DV body axes, initiates and is patterned around the intersection of the two AP and DV compartment boundaries.
The inference of the four constituent compartments provided a powerful framework to interpret the cell behaviors during the subsequent limb bud formation and elongation stages in a quantitative manner. This analysis strongly suggested that a combination of cellular (and associated molecular) mechanisms is at work to transform the twodimensional embryonic epithelium into the three-dimensional outgrowing limb bud. First, there was a significant difference in cell proliferation rates between cells in the center (faster dividing) and the periphery (slower dividing) of the limb primordium from early primordium specification until the global elevation of cells above the level of the epithelium during limb bud formation. Such a "growth-based morphogenesis model" has been the dominant hypothesis for almost 50 years to explain the outgrowth of the vertebrate limb bud (Ede and Law, 1969;Hornbruch and Wolpert, 1970) -although it has been challenged recently (Boehm et al., 2010)) -but has never been implicated as the driving mechanism behind arthropod limb evagination. Second, the observed cell and tissue dynamics raised the question whether formation of the convex surface of the limb bud requires modulation of the basal adhesive properties of the dorsal and ventral cells, similar to the folding of the Drosophila wing bilayer (Fristrom et al., 1993). Third, the outgrowth of the limb was tightly associated -and presumably effected -by two patterned cell behaviors: i) an increased cell proliferation rate at the tip of the limb (at the intersection of the 4 compartments) resembling a putative growth zone which generates many of the new cells necessary for limb outgrowth, especially for the growing distal part of the limb; and ii) a strong bias in the orientation of mitotic divisions parallel to the proximal-distal axis of growth. Although cell proliferation rates were relatively heterogeneous in the rest of the limb, our analysis detected a second region of higher mitotic activity in the cells of the anterior compartment abutting the AP boundary both dorsally and ventrally. Fourth, the different proximal-distal domains of the Parhyale limb could be traced back to distinct mediolateral positions in the early germband stage.
During limb bud formation and elongation, there was a transition and refinement of these positional values along the proximal-distal axis. Fifth, the temporal sequence of segmental subdivisions proceeded as follows (early < late): ischium/merus < basis/ischium < coxa/basis + propodus/dactylus + carpus/propodus < carpus/merus. Sixth, besides the early AP and DV lineage restrictions, we observed a non-heritable proximaldistal separation between cells in the ischium and the merus that were the first limb segments to separate, suggesting that limb segments may represent secondary tissue subdivisions.
Our systems-level approach demonstrates that the comprehensive imaging and fine-scale reconstruction of a developmental process can shed light into functionally interdependent patterning mechanisms operating across multiple scales. We expect that the availability of membrane fluorescent reporters for Parhyale live imaging will shed more light in the future on how cell rearrangements and cell shape changes contribute to limb morphogenesis.

Reconciling genetic with cellular models of limb morphogenesis
As demonstrated in diverse developmental systems, compartment boundaries are important organizers of tissue patterning and growth through the expression of secreted morphogens and other signaling molecules (Dahmann et al., 2011;Lawrence and Struhl, 1996;Matsuda et al., 2016;Restrepo et al., 2014;Tabata and Takei, 2004). In the provided extra independent support for a compartment-based mechanism to pattern the DV axis of Parhyale thoracic limbs. It has been argued that the stripy Wg and Dpp expression observed in the Drosophila leg disc is appropriate for patterning a twodimensional epithelium but not a three-dimensional outgrowing limb (Prpic et al., 2003).
Expression at the tip of three-dimensional embryonic limbs has been demonstrated for Dpp in many studied insect and arthropod species (Angelini and Kaufman, 2005). This is not the case in Parhyale, where we detected a broad expression of Ph-dpp in the lateral ordered Parhyale body plan, we anticipate that the LSFM imaging and tracking approaches described here, together with the recent application of CRSIPR/Cas-based methodologies for genome editing (Kao et al., 2016) will provide excellent material to further explore how morphogens like Dpp regulate form and function at cellular resolution.

. The
PhHS-H2B-mRFPruby-SV40polyA cassette was then excised as an AscI fragment and cloned into the AscI-digested pMinos{3xP3-EGFP} vector (Pavlopoulos et al., 2004), generating plasmid pMi{3xP3-EGFP; PhHS-H2B-mRFPruby}. Three independent transgenic lines were established with this construct for heat-inducible expression of H2B-mRFPruby. The most strongly expressing line was selected for all applications. In this line, nuclear H2B-mRFPruby fluorescence plateaued about 12 hours after heat-shock and high levels of fluorescence persisted for at least 24 hours post heat-shock labeling chromatin in all cells throughout the cell cycle.

Multi-view LSFM imaging of Parhyale embryos
To prepare embryos for LSFM imaging, 2.5-day old transgenic embryos (early germband stage S11; (Browne et al., 2005)) were heat-shocked for 1 hour at 37˚C. About 12 hours later (stage S13), they were mounted individually in a cylinder of 1% low melting agarose (SeaPlaque, Lonza) inside a glass capillary (#701902, Brand GmbH) with their AP axis aligned parallel to the capillary. A 1:4000 dilution of red fluorescent beads (#F-Y050 microspheres, Estapor Merck) were included in the agarose as fiducial markers for multi-view reconstruction. During imaging, the embedded embryo was extruded from the capillary into the chamber filled with artificial seawater supplemented with antibiotics and antimycotics (FASWA; (Kontarakis and Pavlopoulos, 2014)). The FASWA in the chamber was replaced every 12 hours after each heat shock (see below). Embryos were imaged on a Zeiss Lightsheet Z.1 microscope equipped with a 20x/1.0 Plan Apochromat immersion detection objective (at 0.9 zoom) and two 10x/0.2 air illumination objectives producing two light-sheets 5.1 µm thick at the waist and 10.2 µm thick at the edges of a 488 µm x 488 µm field of view.
We started imaging Parhyale embryogenesis from 3 angles/views (the ventral side and the two ventral-lateral sides 45˚ apart from ventral view) during 3 to 4.5 days AEL to avoid damaging the dorsal thin extra-embryonic tissue, and continued imaging from 5 views (adding the two lateral sides 90˚ apart from ventral view) during 4.5 to 8 days AEL. A multi-view acquisition was made every 7.5 min at 26˚C. The H2B-mRFPruby fluorescence levels were replenished regularly every 12 hours by raising the temperature in the chamber from 26˚C to 37˚C and heat-shocking the embryo for 1 hour. Each view (z-stack) was composed of 250 16-bit frames with voxel size 0.254 µm x 0.254 µm x 1 µm. Each 1920 x 1920 pixel frame was acquired using two pivoting light-sheets to achieve a more homogeneous illumination and reduced image distortions caused by light scattering and absorption across the field of view. Each optical slice was acquired with a 561 nm laser and exposure time of 50 msec. With these conditions, Parhyale embryos were recorded routinely for a minimum of 4 days resulting on average in 192 time-points / 240K images / 1.7 TB of raw data per day. for right view), and then resaved as tif files (Schmied et al., 2016).

Multi-view LSFM image processing for 4D reconstruction of
2) Bead-based spatial multi-view registration: In each time-point, each view was aligned to an arbitrary reference view fixed in 3D space (e.g. views 0, 45, 90, 135 aligned to 180) using the bead-based registration option (Preibisch et al., 2010). In each view, fluorescent beads scattered in the agarose were segmented with the Difference-of-Gaussian algorithm using a sigma value of 3 and an intensity threshold of 0.005. Corresponding beads were identified between views and were used to determine the affine transformation model that matched each view to the reference view within each time-point.
3) Fusion by multi-view deconvolution: Spatially registered views were down-sampled twice for time and memory efficient computations during the image fusion step. Input views were then fused into a single output 3D image with a more isotropic resolution using the Fiji plugin for multi-view deconvolution estimated from the point spread functions of the fluorescent beads (Preibisch et al., 2014). The same cropping area containing the entire imaged volume was selected for all time-points. In each time-point, the deconvolved fused image was calculated on GPU in blocks of 256x256x256 pixels with 7 iterations of the Efficient Bayesian method regularized with a Tikhonov parameter of 0.0006.

4) Bead-based temporal registration:
To correct for small drifts of the embryo over the extended imaging periods (e.g. due to agarose instabilities), we stabilized the fused volume over time using the segmented beads (sigma = 1.8 and intensity threshold = 0.005) for temporal registration with the affine transformation model using an all-to-all matching within a sliding window of 5 time-points. 5) Computation of spatiotemporally registered fused volumes: Using the temporal registration parameters, we generated a stabilized time-series of the fused deconvolved 3D images. 6) 4D rendering: The Parhyale embryo was rendered over time from the spatiotemporally registered fused data using Fiji's 3D Viewer.

Lineage reconstruction with the Massive Multi-view Tracker (MaMuT)
MaMuT is a Fiji plugin that can be installed through the Fiji Updater. The source code for MaMuT is available on GitHub (https://github.com/fiji/MaMuT) and detailed tutorials and training datasets can be found at http://imagej.net/MaMuT. In brief, for lineaging purposes, the Parhyale multi-view LSFM raw z-stacks were registered spatiotemporally and the image data together with the registration parameters were converted into the custom HDF5/XML file formats utilized by the BigDataViewer and MaMuT plugins.

29
A new MaMuT annotation project was created for each dataset and appendage of interest.
The nuclei contributing to the T2 limb were identified in the first time-point and tracked manually every 5 time-points except during mitosis, in which case we also tracked one time-point before and one after segregation of the daughter chromosomes during anaphase/telophase. To guarantee the accuracy of our lineage reconstructions, the position of each tracked nucleus was verified in at least two neighboring views and by slicing the data orthogonally in separate Viewer windows. All color-coded cell animations were created with MaMuT's 3D Viewer, which is based on (Schmid et al., 2010).
The graph data structure in MaMuT can handle efficiently up to about a hundred thousand annotations. This number is well within the realm of manually generated annotations, but is normally exceeded by large-scale fully automated lineaging engines like TGMM. For this reason, we also provide users the option to crop the imported TGMM annotation in space and/or in time to make them compatible with MaMuT.

In situ hybridization
In situ hybridizations were carried out as previously described (Rehm et al., 2009). The sequence accession number for Parhyale decapentaplegic (Ph-dpp) is XXX, for Dorsocross (Ph-Doc) XXX, and for engrailed-2 (Ph-en2) XXX. Stained samples were imaged on a Zeiss 880 confocal microscope using the Plan-Apochromat 10x/0.45 and 20x/0.8 objectives. Images were adjusted for brightness/contrast and processed using Fiji (http://fiji.sc) and Photoshop (Adobe Systems Inc). For color overlays, the brightfield image of the Ph-dpp or Ph-Doc BCIP/NBT staining was inverted, false-colored green and merged with the fluorescent signal of the Ph-en2 FastRed staining in magenta and the nuclear DAPI signal in blue.

AUTHOR CONTRIBUTIONS
A.P. and C.W. conceived the project, A.P. and P.T. managed the project, A.P. and B.H. generated the Parhyale imaging data, J.Y.T. and T.P. developed MaMuT, T.P., S.P., P.J.K., P.T. and A.P. performed image analysis, E.S. carried out the in situ hybridizations, A.P. and C.W. analyzed the data and prepared the manuscript with input from all authors. Svendsen, P.C., Ryu, J.R., and Brook, W.J. (2015).  Embryo at mid-germband stage (S13 according to (Browne et al., 2005) (E5a and b rows). Likewise, no cell mixing was detected across the dorsalventral compartment boundary that extended again between the E4b and c rows anteriorly, between cells E4c4-c5, E4d3-d4 and E5a4-a5 distally, and between the E5a and b rows posteriorly.     Rectangles indicate the T2, T3 and T4 limbs shown in Figure 7. Scale bars are 100 μm.