MorphoNet 2.0: An innovative approach for qualitative assessment and segmentation curation of large-scale 3D time-lapse imaging datasets

Recent advances in optical time-lapse microscopy now enable the 3D capture of life’s temporal dynamics, from molecular assemblies to entire organisms (McDole et al., 2018), in cells, embryos, and aggregates. These technologies have thus become crucial for cell and developmental biology. The automated segmentation and tracking of these complex datasets is, however, necessary for their interpretation and quantification (McDole et al., 2018).

Recent progress in the 3D segmentation of animal and plant cells (Pachitariu and Stringer, 2022), or organelles (Cutler et al., 2022) has made this process highly efficient, despite the persistence of algorithm-dependent systematic errors (Liu et al., 2023). The processing of time-lapse datasets, however, remains challenging. The accumulation over hundreds of time points of even a few residual segmentation errors can strongly affect data interpretation and notably disrupt long-term cell tracking. Following the excitement generated by this new generation of imaging systems, which offers the ability to image living systems at unprecedented spatial and temporal scales, a glass ceiling has now been reached due to the quality of these reconstructed data, which fails to scale to real research exploitations.

Beyond the reconstruction requirements of individual research projects, proper metrics are essential for evaluating the performance of bioimage analysis algorithms (Nature Methods, 2024). However, the training and benchmarking of next-generation AI-based segmentation tools fundamentally depend on the availability of accurate ground truth data (Maška et al., 2023). The critical bottleneck is the lack of a massive amount of 3D high-quality reconstructed data to train these systems.

We have decided to take on this challenge and develop a tool to produce high-precision and high-quality 3D datasets for AI. Two major barriers must be overcome to reach this milestone. The first concerns the evaluation of dataset reconstructions: how can we measure the accuracy of segmentations obtained by automated algorithms for these complex 3D and 3D+t datasets? The second focuses on the curation of reconstructed data: how can we correct and significantly enhance segmentations to achieve reliable reconstructions that are essential for major scientific breakthroughs?

2D imaging has long been a cornerstone of biological research, driving significant scientific discoveries with visual validation being largely adequate, as it could be performed directly on a 2D screen. The traditional method involves visually comparing the reconstruction by overlaying it with the acquired data, as seen in classic software like ImageJ (Schneider et al., 2012), Napari (Sofroniew, 2022), or Ilastik (Berg et al., 2019). When discrepancies arise, experts can perform curation and manually adjust the pixel values to align with the desired class (e.g. a cell).

In the context of 3D and 3D+t data, the technique encounters significant obstacles that hinder reliable scientific analysis. The 3D structure of images, inherently unsuited to the 2D digital environment, demands numerous compromises that restrict their interpretability. Projecting 3D objects onto a 2D screen causes data masking, requiring experts to perform extensive manipulations to verify accuracy. Additionally, the high number and density of 3D objects further complicate the process, as visual interference from overlapping objects increases during these adjustments. As a result, voxel-level (3D pixel) curation becomes almost unfeasible using conventional methods, compelling experts to revert to working in 2D. This approach becomes very time-consuming and, therefore, does not allow to obtain a satisfactory reconstruction quality for scientific use. A final obstacle is caused by the much larger size of 3D datasets, which significantly lengthens processing time for each curation action. This computational burden becomes daunting for 3D+t time-lapse datasets, as error propagation between consecutive time points rapidly increases the number of segmentation and object tracking errors needing correction. Restricting the application of advanced image processing tasks to subsets of the image may alleviate this issue.

An unsupervised objective assessment can be derived from a priori knowledge of expected spatial (e.g. smoothness of contours, shape regularity, volume…) or temporal (e.g. stability and smooth evolution, lifetime of objects…) features of the objects (Correia and Pereira, 2002). Features of the object boundaries (e.g. contrast between inside and outside of an object) can also be computed (Correia and Pereira, 2002). More sophisticated strategies have been proposed when no statistically relevant a priori knowledge can be drawn (Valindria et al., 2017). Since curation is typically carried out by expert biologists with limited programming skills, the computation of image features and the projection of relevant metrics onto individual segmented objects should be accessible through user-friendly interfaces. An alternative approach for interacting with 3D segmented images is, therefore, necessary.

We previously developed a novel concept of morphogenetic web browser, MorphoNet (Leggio et al., 2019), to visualize and interact with 3D+t segmented datasets through their meshed representations. This user-friendly web solution required no installation and was suited for datasets of moderate size (up to a thousand cells over a few hundred time points). Since its release, this platform has been used in a variety of morphological studies in plant and animal systems (Manni et al., 2022; Chung et al., 2023) and has been instrumental to interpret the relationship between gene expression and complex evolving cell shapes (Refahi et al., 2021; Guignard et al., 2020; Dardaillon et al., 2020). This web version was restricted to visualizing and interacting with segmented datasets through their precomputed 3D dual meshes. This approach offered significant benefits in terms of online 3D rendering efficiency and data management but it also came with limitations. For example, meshes are simplified representations of the segmented object that may not capture fine details of the original volumetric data and that cannot be validated without comparison to the original raw intensity images. Meshed representations can also make it challenging to perform the precise computations required for detailed geometric and topological analyses (Sophie et al., 2024). Exploration of larger datasets was nonetheless constrained by the computational resource limitations of internet browsers.

We present here MorphoNet 2.0, a major conceptual evolution of the platform, designed to offer powerful tools for the reconstruction, evaluation, and curation of 3D and 3D+t datasets. This innovative approach leverages the richness and redundancy of information embedded within these complex datasets. It addresses the two previously identified challenges: evaluating 3D segmented data without relying on manual ground truth by fully harnessing the data’s richness, and enabling semi-automated curation of 3D data, now achievable with just a few clicks. This enables us to achieve data quality at a level suitable for scientific discovery. We showcase the impact of these advancements by revisiting and enhancing five published animal and plant datasets previously regarded as ground truth (Maška et al., 2023).

We feature a new standalone application running on all major operating systems that exploits the resources of the user’s local machine to explore, segment, assess, and manually curate very large 3D and 3D+t datasets. Like the web version, the MorphoNet application uses the power and versatility of the Unity game engine and includes its main features (Leggio et al., 2019). By overcoming web-based limitations, the application can handle more complex datasets, including heavier segmented voxel images and up to several tens of thousands of objects on a research laptop (Figure 1h, and Supplementary file 1). The standalone application allows users to directly explore private data stored on their own computer, without need for an upload to the MorphoNet server. Subsequent data sharing with other researchers or a wider public in an open science process is facilitated by the MorphoNet server upload functions of the standalone.

Figure 1

Download asset Open asset

MorphoNet is designed with a strong emphasis on user-friendliness, making its use highly intuitive for experimental biologists with no coding experience (See Videos in Materials and Methods). Every feature and interface element has been crafted to ensure a smooth, straightforward user experience, allowing both beginners and experts to utilize its capabilities efficiently. Additionally, the solution is fully open-source allowing bio-image analysis to develop their own features (See Supplementary file 2).

One of the key features of this new standalone application is the automatic integration of a duality between the 3D segmented images and their corresponding meshes (Figure 2). This feature is achieved by linking Python, dedicated to high-performance image processing, with the Unity Game Engine, fully optimized for seamless interaction with the dual meshes. This integration enables access to state-of-the-art reconstruction (Pachitariu and Stringer, 2022; Weigert et al., 2020) tools, including those leveraging AI libraries, while also addressing the challenges of interacting with 3D images by harnessing the powerful interactive capabilities of game engines.

Figure 2

Download asset Open asset

To ensure users can consistently assess segmentation quality, raw intensity images are also available by superimposition. This provides access to the entire workflow, from raw image acquisition to meshed segmentation, allowing simultaneous exploration and visualization of both meshed and voxel images (Figure 1a–f).

We also leveraged Unity’s scene management tools to implement simultaneous visualization of multiple scenes. This feature enables the display of interactive cell lineages in a dedicated, fully connected window, offering essential insights into cellular trajectories within 3D+t datasets.

Use cases

To highlight the efficiency of MorphoNet plugins in detecting and correcting the main types of errors across a broad range of organisms, we present five examples from previously published fluorescent live imaging datasets, representing various levels of quality and complexity. These examples collectively showcase MorphoNet’s powerful segmentation quality assessment and error detection capabilities in different contexts, ranging from low-quality nuclear segmentations to high-quality whole-cell segmentations. While these examples frequently involve the use of Cellpose, the current state-of-the-art tool for 3D dataset segmentation, the focus is not on emphasizing the tool’s power but on demonstrating how advanced tools can be enhanced through integration into MorphoNet’s sophisticated graphical plugin interfaces.

The first use-case introduces two unsupervised metrics for assessing dataset quality in the absence of ground truth on a time-lapse light-sheet microscopy movie of a developing Tribolium castaneum embryo. Using these metrics, we demonstrate the power of an iterative segmentation and training strategy. By applying a Cellpose model, trained on a curated subpopulation, to the entire dataset, we improve segmentation quality.

The second use-case presents unsupervised whole-cell morphological metrics to assess the quality of several independent segmentations of a time-lapse confocal movie of a developing starfish embryo, up to the 512 cell stage, with fluorescently labeled cell plasma membranes. Originally, a bio-image analyst segmented the movie using a complex workflow. We demonstrate that running an imported, custom-trained Cellpose 3D model through MorphoNet’s graphical interfaces greatly increases the number of cells that can be successfully segmented.

The third use-case focuses on detecting and resolving regions of under-segmented cells in a confocal 3D dataset of Arabidopsis thaliana plant apical shoot meristem, with fluorescent plasma membrane and whole-cell segmentation of 1800 cells. It demonstrates the identification of a large region with heavily under-segmented cells, which is successfully addressed by targeting Cellpose to the specific region of interest.

The fourth use-case addresses the correction of low-quality automated cell nuclei segmentation in a time-lapse confocal Caenorhabditis elegans embryo with poor-quality nuclei labeling up to the 350 cell stage, despite accurate manual tracking. It demonstrates how object edition plugins can be used to correct individual nuclei segmentations and propagate these corrections over time.

The final use-case illustrates the efficient detection and correction in the cell lineage of rare residual errors that escaped previous scrutiny in a published time-lapse multiview light-sheet imaging dataset of a developing Phallusia mammillata embryo. The dataset features labeled cell membranes and high-quality whole-cell segmentation and tracking. This polishing work is crucial for generating ground truths for studies of natural variation.

All the evaluation and curation steps for the use cases are detailed in Materials and methods.

Use-case 1: Assessing and improving segmentation quality of Tribolium castaneum embryos using unsupervised nuclei intensity metrics and iterative training on partial annotation

This use-case introduces two unsupervised metrics for assessing dataset quality without ground truth, demonstrating their application in evaluating segmentation performance. It also highlights the effectiveness of an iterative segmentation and training approach, where applying a Cellpose model trained on a curated subset significantly improves overall segmentation quality.

Dataset description

This dataset features a 59-time-step light-sheet microscopy acquisition of labeled cell nuclei from a developing Tribolium castaneum embryo, containing over 7000 cells per time step. It includes two types of ground truth: a gold truth (GT), which is expert manual tracking of 181 nuclei over all time steps, and a silver truth (ST), generated through a complex pipeline combining the outputs of up to 16 top-performing algorithms selected from 39 submissions. The ST provides automated segmentation masks using a modified label fusion approach for improved accuracy in dense datasets. The GT and ST are independent and do not share common elements.

Type of errors

This dataset serves as a benchmark in the fifth Cell Tracking Challenge (Maška et al., 2023) to evaluate and compare AI-based image segmentation methods and is, therefore, regarded as a reliable ground truth. The manual tracking GT is of high quality (expert annotation), but the automated ST is of lesser quality. To evaluate the segmentation quality of the dataset at first time point, the intensity image and the two ground truths (ST and GT) were simultaneously displayed in separate MorphoNet channels (Figure 3b). Systematic visual inspection identified 12/181 GT points without corresponding nucleus in the ST data; eight misplaced segmentations where the GT point is not inside the segmentation; 56 over segmentations and many suspicious nuclear shapes.

Figure 3 with 3 supplements see all

Download asset Open asset

Error identification

Using the Match plugin, each GT position was automatically associated with a corresponding nucleus in the ST, leaving 11 out of 181 positions unassigned (Figure 3b), indicating approximately 6% of missed nuclei in the ST. To identify inaccurately shaped segmentations, we used two signal intensity-based MorphoNet metrics. The first metric measures the distance between the geometric center of a segmented object and the center of mass of the signal intensity (Figure 3—figure supplement 1 and Materials and methods for the full properties description), with 32 of the 56 over-segmented nuclei falling into the upper quartile of this metric’s distribution (Figure 3c and d, Figure 3—figure supplement 2). The second metric evaluates deviations in the signal intensity along the segmentation boundary, with 29 out of 56 over-segmented nuclei identified in the upper quartile (Figure 3c–f, Figure 3—figure supplement 1). These errors were visualized in 3D by projecting both metrics onto each nucleus, enabling spatial identification of potential segmentation issues (Figure 3—figure supplement 2).

Error correction

Due to the high number of errors in the published ST, a de novo nucleus segmentation pipeline was created in MorphoNet with minimal manual intervention. The five-step process (as detailed in Materials and methods UC1 Curated pipeline, Video 1) included initial segmentation using the standard Cellpose nuclei model, manual curation of 181 GT nuclei, iterative training of a custom Cellpose model, and refinement of segmentation using geometric properties. Key steps involved correcting segmentation errors with various plugins, extending the Cellpose nuclei model using curated data, and refining non-convex shapes. The pipeline significantly improved segmentation quality compared to the published ST and standard Cellpose model (Figure 3i). Analysis of the curated nuclei revealed a more heterogeneous volume distribution, fewer over-segmentations, and better alignment of intensity-based metrics (Figure 3g and h, Figure 3—figure supplement 3), showing substantial improvement over the original segmentation pipeline. We also created a new set of manually segmented ground truth cells (see Materials and methods UC1 ground truth protocol), distributed across the ranges of both intensity_offset and intensity_border_variation. Using this reference, we evaluated the published and curated segmentations (Figure 3—figure supplement 2h–j), confirming that our pipeline both reduces segmentation errors and better matches expert annotations. Moreover, the analysis revealed a clear correlation between the unsupervised metrics and segmentation accuracy. These results validate the utility of our metrics and demonstrate that the iterative pipeline yields higher-quality segmentations than the published dataset.

Video 1

Download asset

Use case 2: Evaluating membrane segmentation quality using smoothness metrics and simplifying a segmentation workflow for Patiria miniata starfish embryo

This use case thus introduces new unsupervised morphological metrics (the smoothness property) to assess segmentation quality and demonstrates that using a custom-trained 3D Cellpose model via graphical interfaces allows extending segmentation to areas of the dataset with lower intensity/noise ratio. Visualization of metrics, such as smoothness and cell volumes simplify the identification of cells that need further manual curation.

Dataset description

This dataset comprises 300 3D stacks from time-lapse, two-channel confocal imaging of twelve Patiria miniata wild-type and compressed embryos, captured between the 128- and 512 cell stages (Barone et al., 2024). Fluorescent labeling highlights cell membranes and nuclei, though imaging was limited to about half of each embryo due to poor penetration and the use of low magnification air objectives. The published whole-cell segmentation (see Materials and methods UC2 Original workflow) relied on a complex, multi-step workflow adapted from the CartoCel (Andrés-San Román et al., 2023) requiring advanced bioanalysis expertise. For this use case, a particularly challenging dataset - a compressed wild-type embryo at the 512 cell stage - was selected for evaluation and curation.

MorphoNet segmentation with the advanced cellpose plugin

We explored cellpose’s potential to improve segmentation quality in challenging regions of a 3D dataset. Initial segmentation using the Cyto2 model identified 1187 cells, with a bimodal size distribution and 939 cells smaller than 1000 voxels (Figure 4i). The Deli plugin, which eliminates these small cells (Figure 4j), brought the number of cells down to 211 (Figure 4j); however, analysis with the smoothness property (Figure 4g) revealed rough cell surfaces. To address this, we further extend the training of the Cyto2 model on the published 3D database (Barone et al., 2024) (see Materials and methods UC2 Training Protocol workflow, Video 2). The retrained model produced 284 cells with, again, a bimodal size distribution (Figure 4i) which was corrected with the Deli plugin. This significantly simpler pipeline produced smooth cell segmentations (Figure 4h), with a size distribution comparable to the published ground truth and a high Intersection over Union (IoU) score (Figure 4—figure supplement 1). While 2 cells were missing, and 11 under-segmented compared to the published data (Figure 4f), this new approach recovered 48 additional cells (Figure 4e) while simplifying the segmentation process compared to the original pipeline.

Figure 4 with 1 supplement see all

Download asset Open asset

Video 2

Download asset

Use case 3: Interactive targeted segmentation using cellpose for resolving under-segmentation in Arabidopsis thaliana shoot apical meristem whole-cell segmentation

This use case demonstrates the power of interactive selection with targeted segmentation methods that can resolve under-segmentation issues in complex datasets.

Dataset description

This dataset (Willis et al., 2016) consists of a 19-time steps time-lapse confocal acquisition of a live Arabidopsis thaliana shoot apical meristem with fluorescently-labeled cell membranes. Cell numbers ranged from 600 cells at the first time point to around 1800 cells at the last (Figure 5a). We use the last time step (19) with the higher number of cells to illustrate the curation procedure.

Figure 5 with 1 supplement see all

Download asset Open asset

Error detection

The analysis of the published segmentation uncovered multiple under- and over-segmentation errors in the deep cell layers, caused by poor image quality in the inner regions of the meristem (Figure 5b). These errors hindered accurate cell tracking over time. Observing the size homogeneity of shoot apical meristem cells (Figure 5d), we hypothesized that larger-than-expected cells might indicate under-segmentation errors, while smaller cells could result from over-segmentation. By projecting the automatically computed cell volumes as a color map onto the segmentations, we identified a large under-segmented region containing 20 cells in the deep meristem layer (highlighted in red in Figure 5c), along with a few small over-segmented cells (Figure 5i).

Error correction

The large number of fused cells in the under-segmented region prompted us to test whether our interactive Cellpose plugin could outperform the original MARS/ALT software (Fernandez et al., 2010). The standard pretrained cyto2 model of Cellpose (Pachitariu and Stringer, 2022) produced numerous very small over-segmented cells (Figure 5d, Figure 5—figure supplement 1), suggesting it was not well-suited for this dataset (Kar et al., 2022). To improve performance, we used the high-quality MARS/ALT segmentations from the first ten time points and employed our Cellpose train plugin to extend the cyto2 model training. We tested two training modalities: ‘2D’ training, using only the XY planes, and ‘3D’ training, incorporating XY, XZ, and YZ planes. The segmentations after additional 2D training still contained many over-segmented cells (Figure 5f), a problem that was significantly reduced with 3D training (Figure 5g). Despite this, the heterogeneity of signal intensities still caused over-segmentation in more superficial regions of the meristem. We thus targeted the Cellpose plugin only to the large under-segmented region (Figure 5h) and removed small over-segmented cells using the Deli segmentation correction plugin (Figure 5i). This targeted approach was fast and accurate (Figure 5d) and allowed us to preserve the accurate segmentations from the original published data while curating the identified errors. Notably, the curated result restored a unimodal and symmetric volume distribution (Figure 5—figure supplement 1g), consistent with the expected homogeneity of the shoot apical meristem (Shapiro et al., 2015), which was further confirmed by visual inspection (Figure 5—figure supplement 1h–i). With this approach, we successfully curated 20 under-segmented regions and generated 98 new cells (Video 3).

Video 3

Download asset

Use-case 4: Improving segmentation quality using object editing plugins for Caenorhabditis elegans cell lineage

This use case illustrates how object editing plugins can be employed to improve segmentation quality in challenging datasets despite poor-quality nuclei labeling.

Dataset description

This dataset (Murray et al., 2008) consists of a live 3D time-lapse confocal acquisition of a developing Caenorhabditis elegans embryo with fluorescently labeled cell nuclei (Figure 6a), over 195 time steps between the 4- and 362 cell stages. This dataset was included in the Cell Tracking Challenge (Maška et al., 2023) (see Materials and methods UC4 for the original workflow).

Figure 6 with 1 supplement see all

Download asset Open asset

Error detection

Although the manually curated published cell lineage appeared accurate (Figure 6a), some touching nuclei were elongated with flat vertical contact interfaces, likely due to poor imaging quality. To identify these segmentation artifacts systematically, we calculated the axis ratio for each segmented object and visualized these values directly on their surface meshes (Figure 6b). The distribution revealed two peaks: a sharp one around 2, containing most nuclei, and a broader one above 3. Nuclei with an axis ratio greater than 2.9 were flagged for potential correction.

Error correction

We curated these errors for a specific time point by first applying a fusion correction plugin (Figure 6f and g) to all selected nuclei pairs with an axis ratio greater than 2.9 (Figure 6e). This was followed by a separation plugin (Figure 6h) that divides fused objects into distinct nuclei using a Gaussian mixture applied to intensity values. We initially addressed the more complex cases manually, which included one group of six fused nuclei, one group of four fused nuclei, and two groups of three. Next, we automatically curated the remaining nine simpler cases of two fused nuclei by selecting nuclei with a volume greater than 80 µm³. As expected, the axis ratio distribution of the curated set became unimodal and centered around 2 (Figure 6b), reflecting more regular nuclear shapes—an improvement confirmed by visual inspection (Figure 6—figure supplement 1). The entire time step at t=150 was fully curated with just 5 plugin actions (Figure 6h, Video 4).

Video 4

Download asset

Use-case 5: Enhancing cell lineage accuracy by detecting segmentation errors in Phallusia mammillata embryos

This use case showcases how MorphoNet can efficiently detect and correct segmentation errors in complex 3D+t datasets of developing ascidian embryos, enhancing the accuracy of cell lineage reconstructions and analysis of cell division timing variations.

Dataset description

This dataset (Guignard et al., 2020) comprises 64 time steps of multi-view light-sheet microscopy of developing ascidian embryos (64–298 cell stages), segmented using the ASTEC algorithm (Guignard et al., 2020). Despite the high-quality segmentation and tracking, residual errors, such as multi-component cells, over- and under-segmentations, delayed, or missed divisions, and shape inaccuracies remained, with no tools available at the time of the publication to systematically validate or correct the 10,000 cell snapshots.

Error detection

MorphoNet was used to polish the cell lineage to a sufficient level to study natural variation in cell division timing. The main challenge was identifying rare errors within a cell lineage that appeared accurate by visual inspection. To address common lineage residual errors, such as missing or delayed divisions and broken lineage links, we developed a set of segmentation error identification metrics. These metrics were visualized by projecting their results onto the lineage using an interactive viewer linked to the 3D dataset (Figure 7a and b). All identified errors stemmed from segmentation issues rather than tracking, and their correction was streamlined by the bidirectional connection between the lineage and embryo representation windows, allowing direct navigation between the two.

Figure 7 with 1 supplement see all

Download asset Open asset

State-of-the-art segmentation methods often produce systematic errors, such as very small segmented objects, which we identified using geometric properties like cell volume (e.g. cell a7.1, Figure 7a). Property projection, combined with scatter visualization, facilitates the identification of cells even in dense, interior regions (Figure 7b), while the highlight mode enhances this process by isolating the cell of interest (Figure 7c). Occasional missing or false divisions disrupt accurate cell histories. During early ascidian embryogenesis, stable cell volumes allowed volume projections onto the lineage to identify rapid variations, revealing missed divisions (Figure 7d).

Ascidian embryos exhibit bilateral symmetry, with homologous cells dividing almost synchronously. A lineage distance property (Guignard et al., 2020) revealed missed divisions (e.g. b7.11* in Figure 7e). To distinguish segmentation errors from natural variability, a compactness property was used to detect inaccuracies in division timing by tracking cell rounding during mitosis (e.g. Figure 7f).

Using MorphoNet, an expert identified approximately 20 isolated cell lineage errors in a dataset of 185 cell snapshots in just 2 hr.

Error correction

The 41 multi-component labels were separated using the Disco plugin, identifying seven potential cells via the volume property analysis (Figure 7b). The others, considered as small artifacts, were merged with neighboring cells sharing the largest surface area using the Deli plugin. Division timing issues often stem from over-segmentation, under-segmentation, or missing cells. Over-segmentations were resolved with the Fuse plugin (Figure 7g), while under-segmentations were corrected using Propagate segmentation plugins by tracing back from the first accurate segmentation of sister cells to their mother’s division point (Figure 7h). For missing cells, seeds were added with seed generator plugins and segmented using local seeded watershed algorithms (Figure 7i). If imaging quality was too poor, the bilaterally symmetrical cell served as a mirror-image proxy, using the Copy-Paste plugin to replicate, rotate, and scale the cell to the missing position (Figure 7j). A total of 185 errors were corrected in approximately eight hours using 264 MorphoNet plugin actions, resulting in a more accurate estimation of natural variability in cell division timing (Figure 7k, Figure 7—figure supplement 1). This duration includes the full workflow performed by a single user: loading the dataset, exploring the cell lineage, identifying segmentation or division errors, and applying corrective plugins. The reported actions include both exploratory attempts and validated corrections, combining user interaction with computation time (Video 5).

Video 5

Download asset

Recent advancements in optical time-lapse microscopy allow for the 3D capture of dynamic biological processes but face challenges in automating the segmentation and tracking of large, complex datasets. Residual segmentation errors in time-lapse datasets disrupt data interpretation and hinder long-term cell tracking. Moreover, accurate, high-quality 3D data is critical for training next-generation AI-based segmentation tools, yet the availability of such datasets remains limited.

To address these challenges, MorphoNet 2.0 was developed as a standalone application for reconstructing, evaluating, and curating 3D and 3D+t datasets. Building on its previous web-based version, MorphoNet 2.0 integrates voxel images with meshed representations, leveraging both Unity Game Engine and Python for enhanced interactivity and processing power. Key features include tools for assessing segmentation quality through unsupervised metrics (e.g., volume, smoothness, and temporal stability), automated error detection, and visualization of segmented objects alongside raw intensity images.

Conventional image curation tools struggle with the complexities of interacting with 3D voxel images. In contrast, MorphoNet 2.0 introduces a novel approach using dual representation layers, enabling efficient biocuration by isolating problem areas and significantly accelerating dataset refinement. MorphoNet 2.0 is user-friendly, open-source, and supports both scientific discovery and AI training by producing high-precision 3D datasets and enabling reproducible, scalable data analysis.

By showcasing five use-cases of fluorescence datasets in which cell membranes or nuclei were labeled, we demonstrated that the tool’s high versatility and user-friendliness enables biologists without programming skills to efficiently and intuitively detect and handle a broad range of errors (under-segmentation, over-segmentation, missing objects, lineage errors).

MorphoNet has a high potential to adapt to evolving datasets and segmentation challenges. First, its open-source Python plugin architecture fosters community-driven improvements. These could target cellular datasets as exemplified by the five use-cases presented. They could also open MorphoNet to other imaging modalities, including multi-modal datasets combining, for instance, fluorescence and electron microscopy. Additional plugins could accommodate new AI tools or automate the training of segmentation models through data augmentation, feature selection, or hyperparameter optimization. Plugins for widely used platforms, such as Napari or Fiji will also broaden the user-base and interoperability of the tool, as would also the development of flexible export options to integrate MorphoNet outputs with other analytical pipelines or visualization software. The MorphoNet platform could be further extended through the creation of a centralized repository for community-developed plugins, the organization of MorphoNet-based bio-image analysis challenges to stimulate community engagement, and the provision of curated datasets to serve as benchmarks for testing and validating new segmentation algorithms. To support these developments over time, we rely on institutional support from our host laboratories and research organizations, and we will seek additional funding through dedicated research grants. We also aim to foster open-source contributions, develop training materials, and provide user support to ensure long-term adoption and sustainability.

Curation could also be improved by the introduction of cloud-based, multi-user capabilities to enable experts to work on the same dataset simultaneously. For now, web browsers impose constraints on the use of computing resources, which could be lifted in the near future, for example, through the development of WebGPU.

Curation capabilities will likely also be enhanced by the implementation of adaptive machine learning models that integrate past user corrections to suggest or automate future edits. In the context of 3D segmentation tasks, manual expertise and curation for voxel-wise segmentation is labor-intensive and expensive. Full annotation of large datasets may thus not be feasible. Partial annotations allow datasets to be created with less time and resources while still providing valuable information. Including algorithms, such as Sketchpose (Cazorla et al., 2025) could leverage weakly-supervised learning to generalize from the partially annotated data and infer segmentation patterns in the unlabeled portions of the dataset, a strategy we initiated with the Tribolium dataset. This will make it possible to train models on diverse datasets without the burden of full annotations.

By addressing major challenges in 3D and 3D+t dataset assessment and curation, MorphoNet 2.0 provides a versatile platform for improving segmentation quality and generating reliable ground truths. Its user-friendliness, adaptability, and extensibility position it as a valuable tool for advancing quantitative bio-image analysis, with significant potential for enhancement and broader application in the future.

The MorphoNet_Data repository (https://doi.org/10.6084/m9.figshare.30529745.v1) provides all datasets and resources associated with the MorphoNet 2.0 publication. It serves as an open archive supporting the reproducibility and reuse of the study's results. The repository includes: DATASETS - Original and curated imaging datasets used in MorphoNet 2.0. BENCHMARKING_DATASETS - Reference datasets employed for performance evaluation, as detailed in the benchmark documentation and supplementary materials. MODELS - Pre-trained Cellpose models specifically used for segmentation curation in each dataset. MOVIES - Demonstration videos illustrating the use cases of each dataset.

1. 1. Gallean B
   2. Laurent T
   3. Biasuz K
   4. Clement A
   5. Faraj N
   6. Lemaire P
   7. Faure E

   (2025) figshare

   MorphoNet 2.0: An innovative approach for qualitative assessment and segmentation curation of large-scale 3D time-lapse imaging datasets.

   https://doi.org/10.6084/m9.figshare.30529745.v1

1. Conference

   1. Abadi M

   (2016)

   TensorFlow: A system for large-scale machine learning

   12th USENIX1468 Symposium on Operating Systems Design and Implementation (OSDI 16). pp. 265–283.

   • Google Scholar
2. 1. Andrés-San Román JA
   2. Gordillo-Vázquez C
   3. Franco-Barranco D
   4. Morato L
   5. Fernández-Espartero CH
   6. Baonza G
   7. Tagua A
   8. Vicente-Munuera P
   9. Palacios AM
   10. Gavilán MP
   11. Martín-Belmonte F
   12. Annese V
   13. Gómez-Gálvez P
   14. Arganda-Carreras I
   15. Escudero LM

   (2023) CartoCell, a high-content pipeline for 3D image analysis, unveils cell morphology patterns in epithelia

   Cell Reports Methods 3:100597.

   https://doi.org/10.1016/j.crmeth.2023.100597

   • PubMed
   • Google Scholar
3. 1. Barone V
   2. Tagua A
   3. Román J
   4. Hamdoun A
   5. Garrido-García J
   6. Lyons DC
   7. Escudero LM

   (2024) Local and global changes in cell density induce reorganisation of 3D packing in a proliferating epithelium

   Development 151:dev202362.

   https://doi.org/10.1242/dev.202362

   • PubMed
   • Google Scholar
4. 1. Berg S
   2. Kutra D
   3. Kroeger T
   4. Straehle CN
   5. Kausler BX
   6. Haubold C
   7. Schiegg M
   8. Ales J
   9. Beier T
   10. Rudy M
   11. Eren K
   12. Cervantes JI
   13. Xu B
   14. Beuttenmueller F
   15. Wolny A
   16. Zhang C
   17. Koethe U
   18. Hamprecht FA
   19. Kreshuk A

   (2019) ilastik: interactive machine learning for (bio)image analysis

   Nature Methods 16:1226–1232.

   https://doi.org/10.1038/s41592-019-0582-9

   • PubMed
   • Google Scholar
5. 1. Boyle TJ
   2. Bao Z
   3. Murray JI
   4. Araya CL
   5. Waterston RH

   (2006) AceTree: a tool for visual analysis of Caenorhabditis elegans embryogenesis

   BMC Bioinformatics 7:275.

   https://doi.org/10.1186/1471-2105-7-275

   • PubMed
   • Google Scholar
6. 1. Cazorla C
   2. Munier N
   3. Morin R
   4. Weiss P

   (2025) Sketchpose: learning to segment cells with partial annotations

   Machine Learning for Biomedical Imaging 3:367–381.

   https://doi.org/10.59275/j.melba.2025-f7b3

   • Google Scholar
7. 1. Child CM

   (1906) The organization and cell-lineage of the ascidian egg

   Science 23:340–344.

   https://doi.org/10.1126/science.23.583.340

   • Google Scholar
8. 1. Chung J
   2. Newman-Smith E
   3. Kourakis MJ
   4. Miao Y
   5. Borba C
   6. Medina J
   7. Laurent T
   8. Gallean B
   9. Faure E
   10. Smith WC

   (2023) A single oscillating proto-hypothalamic neuron gates taxis behavior in the primitive chordate Ciona

   Current Biology 33:3360–3370.

   https://doi.org/10.1016/j.cub.2023.06.080

   • PubMed
   • Google Scholar
9. 1. Correia PL
   2. Pereira F

   (2002) Stand-alone objective segmentation quality evaluation

   EURASIP Journal on Advances in Signal Processing 2002:1–12.

   https://doi.org/10.1155/S1110865702000707

   • Google Scholar
10. 1. Cutler KJ
    2. Stringer C
    3. Lo TW
    4. Rappez L
    5. Stroustrup N
    6. Brook Peterson S
    7. Wiggins PA
    8. Mougous JD

    (2022) Omnipose: a high-precision morphology-independent solution for bacterial cell segmentation

    Nature Methods 19:1438–1448.

    https://doi.org/10.1038/s41592-022-01639-4

    • PubMed
    • Google Scholar
11. 1. Dardaillon J
    2. Dauga D
    3. Simion P
    4. Faure E
    5. Onuma TA
    6. DeBiasse MB
    7. Louis A
    8. Nitta KR
    9. Naville M
    10. Besnardeau L
    11. Reeves W
    12. Wang K
    13. Fagotto M
    14. Guéroult-Bellone M
    15. Fujiwara S
    16. Dumollard R
    17. Veeman M
    18. Volff J-N
    19. Roest Crollius H
    20. Douzery E
    21. Ryan JF
    22. Davidson B
    23. Nishida H
    24. Dantec C
    25. Lemaire P

    (2020) ANISEED 2019: 4D exploration of genetic data for an extended range of tunicates

    Nucleic Acids Research 48:D668–D675.

    https://doi.org/10.1093/nar/gkz955

    • PubMed
    • Google Scholar
12. 1. Fernandez R
    2. Das P
    3. Mirabet V
    4. Moscardi E
    5. Traas J
    6. Verdeil J-L
    7. Malandain G
    8. Godin C

    (2010) Imaging plant growth in 4D: robust tissue reconstruction and lineaging at cell resolution

    Nature Methods 7:547–553.

    https://doi.org/10.1038/nmeth.1472

    • Google Scholar
13. 1. Franco-Barranco D
    2. Pastor-Tronch J
    3. González-Marfil A
    4. Muñoz-Barrutia A
    5. Arganda-Carreras I

    (2022) Deep learning based domain adaptation for mitochondria segmentation on EM volumes

    Computer Methods and Programs in Biomedicine 222:106949.

    https://doi.org/10.1016/j.cmpb.2022.106949

    • PubMed
    • Google Scholar
14. 1. Guignard L
    2. Fiúza U-M
    3. Leggio B
    4. Laussu J
    5. Faure E
    6. Michelin G
    7. Biasuz K
    8. Hufnagel L
    9. Malandain G
    10. Godin C
    11. Lemaire P

    (2020) Contact area-dependent cell communication and the morphological invariance of ascidian embryogenesis

    Science 369:eaar5663.

    https://doi.org/10.1126/science.aar5663

    • PubMed
    • Google Scholar
15. 1. Harris CR
    2. Millman KJ
    3. van der Walt SJ
    4. Gommers R
    5. Virtanen P
    6. Cournapeau D
    7. Wieser E
    8. Taylor J
    9. Berg S
    10. Smith NJ
    11. Kern R
    12. Picus M
    13. Hoyer S
    14. van Kerkwijk MH
    15. Brett M
    16. Haldane A
    17. Del Río JF
    18. Wiebe M
    19. Peterson P
    20. Gérard-Marchant P
    21. Sheppard K
    22. Reddy T
    23. Weckesser W
    24. Abbasi H
    25. Gohlke C
    26. Oliphant TE

    (2020) Array programming with NumPy

    Nature 585:357–362.

    https://doi.org/10.1038/s41586-020-2649-2

    • PubMed
    • Google Scholar
16. 1. Kar A
    2. Petit M
    3. Refahi Y
    4. Cerutti G
    5. Godin C
    6. Traas J

    (2022) Benchmarking of deep learning algorithms for 3D instance segmentation of confocal image datasets

    PLOS Computational Biology 18:e1009879.

    https://doi.org/10.1371/journal.pcbi.1009879

    • PubMed
    • Google Scholar
17. Software

    1. Laurent T

    (2025a) MorphoNet_Unity, version 61010d9f

    GitLab.

    https://gitlab.inria.fr/MorphoNet/morphonet_unity
18. Software

    1. Laurent T

    (2025b) MorphoNet_API, version a1f72212

    GitLab.

    https://gitlab.inria.fr/MorphoNet/morphonet_api
19. Software

    1. Laurent T

    (2025c) MorphoNet_Lineage, version b784523c

    GitLab.

    https://gitlab.inria.fr/MorphoNet/morphonet_lineage
20. 1. Leggio B
    2. Laussu J
    3. Carlier A
    4. Godin C
    5. Lemaire P
    6. Faure E

    (2019) MorphoNet: an interactive online morphological browser to explore complex multi-scale data

    Nature Communications 10:2812.

    https://doi.org/10.1038/s41467-019-10668-1

    • PubMed
    • Google Scholar
21. 1. Liu Y
    2. Jin Y
    3. Azizi E
    4. Blumberg AJ

    (2023) Cellstitch: 3D cellular anisotropic image segmentation via optimal transport

    BMC Bioinformatics 24:480.

    https://doi.org/10.1186/s12859-023-05608-2

    • PubMed
    • Google Scholar
22. 1. Manni L
    2. Caicci F
    3. Anselmi C
    4. Vanni V
    5. Mercurio S
    6. Pennati R

    (2022) Morphological study and 3D reconstruction of the larva of the ascidian halocynthia roretzi

    Journal of Marine Science and Engineering 10:11.

    https://doi.org/10.3390/jmse10010011

    • Google Scholar
23. 1. Maška M
    2. Ulman V
    3. Delgado-Rodriguez P
    4. Gómez-de-Mariscal E
    5. Nečasová T
    6. Guerrero Peña FA
    7. Ren TI
    8. Meyerowitz EM
    9. Scherr T
    10. Löffler K
    11. Mikut R
    12. Guo T
    13. Wang Y
    14. Allebach JP
    15. Bao R
    16. Al-Shakarji NM
    17. Rahmon G
    18. Toubal IE
    19. Palaniappan K
    20. Lux F
    21. Matula P
    22. Sugawara K
    23. Magnusson KEG
    24. Aho L
    25. Cohen AR
    26. Arbelle A
    27. Ben-Haim T
    28. Raviv TR
    29. Isensee F
    30. Jäger PF
    31. Maier-Hein KH
    32. Zhu Y
    33. Ederra C
    34. Urbiola A
    35. Meijering E
    36. Cunha A
    37. Muñoz-Barrutia A
    38. Kozubek M
    39. Ortiz-de-Solórzano C

    (2023) The cell tracking challenge: 10 years of objective benchmarking

    Nature Methods 20:1010–1020.

    https://doi.org/10.1038/s41592-023-01879-y

    • PubMed
    • Google Scholar
24. 1. McDole K
    2. Guignard L
    3. Amat F
    4. Berger A
    5. Malandain G
    6. Royer LA
    7. Turaga SC
    8. Branson K
    9. Keller PJ

    (2018) In toto imaging and reconstruction of post-implantation mouse development at the single-cell level

    Cell 175:859–876.

    https://doi.org/10.1016/j.cell.2018.09.031

    • PubMed
    • Google Scholar
25. 1. Murray JI
    2. Bao Z
    3. Boyle TJ
    4. Waterston RH

    (2006) The lineaging of fluorescently-labeled Caenorhabditis elegans embryos with StarryNite and AceTree

    Nature Protocols 1:1468–1476.

    https://doi.org/10.1038/nprot.2006.222

    • PubMed
    • Google Scholar
26. 1. Murray JI
    2. Bao Z
    3. Boyle TJ
    4. Boeck ME
    5. Mericle BL
    6. Nicholas TJ
    7. Zhao Z
    8. Sandel MJ
    9. Waterston RH

    (2008) Automated analysis of embryonic gene expression with cellular resolution in C. elegans

    Nature Methods 5:703–709.

    https://doi.org/10.1038/nmeth.1228

    • PubMed
    • Google Scholar
27. 1. Najman L
    2. Schmitt M

    (1994) Watershed of a continuous function

    Signal Processing 38:99–112.

    https://doi.org/10.1016/0165-1684(94)90059-0

    • Google Scholar
28. 1. Nature Methods

    (2024) Where imaging and metrics meet

    Nature Methods 21:151.

    https://doi.org/10.1038/s41592-024-02187-9

    • Google Scholar
29. 1. Pachitariu M
    2. Stringer C

    (2022) Cellpose 2.0: how to train your own model

    Nature Methods 19:1634–1641.

    https://doi.org/10.1038/s41592-022-01663-4

    • PubMed
    • Google Scholar
30. Preprint

    1. Paszke A

    (2019) PyTorch: an imperative style, high-performance deep learning library

    arXiv.

    https://arxiv.org/abs/1912.01703

    • Google Scholar
31. Software

    1. Pedregosa F

    (2025)

    Scikit-learn: machine learning in python

    Mach Learn PYTHON.
32. Book

    1. Pinidiyaarachchi A
    2. Wählby C

    (2006) ICIAP

    In: Roli F, Vitulano S, editors. Seeded Watersheds for Combined Segmentation and Tracking of Cells. Springer. pp. 336–343.

    https://doi.org/10.1007/11553595_41

    • Google Scholar
33. 1. Refahi Y
    2. Zardilis A
    3. Michelin G
    4. Wightman R
    5. Leggio B
    6. Legrand J
    7. Faure E
    8. Vachez L
    9. Armezzani A
    10. Risson AE
    11. Zhao F
    12. Das P
    13. Prunet N
    14. Meyerowitz EM
    15. Godin C
    16. Malandain G
    17. Jönsson H
    18. Traas J

    (2021) A multiscale analysis of early flower development in Arabidopsis provides an integrated view of molecular regulation and growth control

    Developmental Cell 56:540–556.

    https://doi.org/10.1016/j.devcel.2021.01.019

    • PubMed
    • Google Scholar
34. 1. Schneider CA
    2. Rasband WS
    3. Eliceiri KW

    (2012) NIH Image to ImageJ: 25 years of image analysis

    Nature Methods 9:671–675.

    https://doi.org/10.1038/nmeth.2089

    • PubMed
    • Google Scholar
35. 1. Shapiro BE
    2. Tobin C
    3. Mjolsness E
    4. Meyerowitz EM

    (2015) Analysis of cell division patterns in the Arabidopsis shoot apical meristem

    PNAS 112:4815–4820.

    https://doi.org/10.1073/pnas.1502588112

    • PubMed
    • Google Scholar
36. Software

    1. Sofroniew N

    (2022) Napari: a multi-dimensional image viewer for python

    Zenodo.

    https://doi.org/10.5281/zenodo.7276432
37. Preprint

    1. Sophie T
    2. Mendieta-Serrano MA
    3. Chapa-y-Lazo, Juliet Chen B
    4. Saunders TE

    (2024) CellMet: extracting 3D shape metrics from cells and tissues

    bioRxiv.

    https://doi.org/10.1101/2024.10.11.617843v1

    • Google Scholar
38. 1. Stegmaier J
    2. Amat F
    3. Lemon WC
    4. McDole K
    5. Wan Y
    6. Teodoro G
    7. Mikut R
    8. Keller PJ

    (2016) Real-time three-dimensional cell segmentation in large-scale microscopy data of developing embryos

    Developmental Cell 36:225–240.

    https://doi.org/10.1016/j.devcel.2015.12.028

    • PubMed
    • Google Scholar
39. 1. Stringer C
    2. Wang T
    3. Michaelos M
    4. Pachitariu M

    (2021) Cellpose: a generalist algorithm for cellular segmentation

    Nature Methods 18:100–106.

    https://doi.org/10.1038/s41592-020-01018-x

    • PubMed
    • Google Scholar
40. 1. Valindria VV
    2. Lavdas I
    3. Bai W
    4. Kamnitsas K
    5. Aboagye EO
    6. Rockall AG
    7. Rueckert D
    8. Glocker B

    (2017) Reverse classification accuracy: predicting segmentation performance in the absence of ground truth

    IEEE Transactions on Medical Imaging 36:1597–1606.

    https://doi.org/10.1109/TMI.2017.2665165

    • PubMed
    • Google Scholar
41. 1. van der Walt S
    2. Schönberger JL
    3. Nunez-Iglesias J
    4. Boulogne F
    5. Warner JD
    6. Yager N
    7. Gouillart E
    8. Yu T
    9. scikit-image contributors

    (2014) scikit-image: image processing in Python

    PeerJ 2:e453.

    https://doi.org/10.7717/peerj.453

    • PubMed
    • Google Scholar
42. 1. Vergara HM
    2. Pape C
    3. Meechan KI
    4. Zinchenko V
    5. Genoud C
    6. Wanner AA
    7. Mutemi KN
    8. Titze B
    9. Templin RM
    10. Bertucci PY
    11. Simakov O
    12. Dürichen W
    13. Machado P
    14. Savage EL
    15. Schermelleh L
    16. Schwab Y
    17. Friedrich RW
    18. Kreshuk A
    19. Tischer C
    20. Arendt D

    (2021) Whole-body integration of gene expression and single-cell morphology

    Cell 184:4819–4837.

    https://doi.org/10.1016/j.cell.2021.07.017

    • PubMed
    • Google Scholar
43. Conference

    1. Weigert M
    2. Schmidt U
    3. Haase R
    4. Sugawara K
    5. Myers G

    (2020) Star-convex Polyhedra for 3D Object Detection and Segmentation in Microscopy

    2020 IEEE Winter Conference on Applications of Computer Vision (WACV).

    https://doi.org/10.1109/WACV45572.2020.9093435

    • Google Scholar
44. 1. Willis L
    2. Refahi Y
    3. Wightman R
    4. Landrein B
    5. Teles J
    6. Huang KC
    7. Meyerowitz EM
    8. Jönsson H

    (2016) Cell size and growth regulation in the Arabidopsis thaliana apical stem cell niche

    PNAS 113:E8238–E8246.

    https://doi.org/10.1073/pnas.1616768113

    • PubMed
    • Google Scholar
45. 1. Wolny A
    2. Cerrone L
    3. Vijayan A
    4. Tofanelli R
    5. Barro AV
    6. Louveaux M
    7. Wenzl C
    8. Strauss S
    9. Wilson-Sánchez D
    10. Lymbouridou R
    11. Steigleder SS
    12. Pape C
    13. Bailoni A
    14. Duran-Nebreda S
    15. Bassel GW
    16. Lohmann JU
    17. Tsiantis M
    18. Hamprecht FA
    19. Schneitz K
    20. Maizel A
    21. Kreshuk A

    (2020) Accurate and versatile 3D segmentation of plant tissues at cellular resolution

    eLife 9:e57613.

    https://doi.org/10.7554/eLife.57613

    • PubMed
    • Google Scholar

Author details

1. Benjamin Gallean

   1. Laboratoire d'informatique, de robotique et de microélectronique de Montpellier, LIRMM, Université de Montpellier, CNRS, Montpellier, France
   2. Centre de Recherche de Biologie cellulaire de Montpellier, CRBM, Université de Montpellier, CNRS, Montpellier, France
   3. Montpellier Ressources Imagerie, Biocampus, Université de Montpellier, CNRS, INSERM, Montpellier, France

   Contribution

   Conceptualization, Data curation, Software, Visualization

   Contributed equally with

   Tao Laurent

   Competing interests

   No competing interests declared

2. Tao Laurent

   Laboratoire d'informatique, de robotique et de microélectronique de Montpellier, LIRMM, Université de Montpellier, CNRS, Montpellier, France

   Contribution

   Conceptualization, Data curation, Software, Visualization

   Contributed equally with

   Benjamin Gallean

   Competing interests

   No competing interests declared

3. Kilian Biasuz

   Centre de Recherche de Biologie cellulaire de Montpellier, CRBM, Université de Montpellier, CNRS, Montpellier, France

   Contribution

   Conceptualization, Data curation, Validation, Visualization

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-0240-2754

4. Ange Clement

   Laboratoire d'informatique, de robotique et de microélectronique de Montpellier, LIRMM, Université de Montpellier, CNRS, Montpellier, France

   Contribution

   Software

   Competing interests

   No competing interests declared

5. Noura Faraj

   Laboratoire d'informatique, de robotique et de microélectronique de Montpellier, LIRMM, Université de Montpellier, CNRS, Montpellier, France

   Contribution

   Software

   Competing interests

   No competing interests declared

6. Patrick Lemaire

   Centre de Recherche de Biologie cellulaire de Montpellier, CRBM, Université de Montpellier, CNRS, Montpellier, France

   Contribution

   Conceptualization, Supervision, Investigation, Methodology, Writing – review and editing

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0003-4925-2009

7. Emmanuel Faure

   Laboratoire d'informatique, de robotique et de microélectronique de Montpellier, LIRMM, Université de Montpellier, CNRS, Montpellier, France

   Contribution

   Conceptualization, Resources, Data curation, Software, Formal analysis, Supervision, Funding acquisition, Validation, Investigation, Visualization, Methodology, Writing – original draft, Project administration, Writing – review and editing

   For correspondence

   emmanuel.faure@lirmm.fr

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0003-2787-0885

• 534

  views

• 19

  downloads

• 1

  citation

Views, downloads and citations are aggregated across all versions of this paper published by eLife.

Citations by DOI

• 1

  citation for Reviewed Preprint v1 https://doi.org/10.7554/eLife.106227.1