Large-volume fully automated cell reconstruction generates a cell atlas of plant tissues

Abstract

The geometric shape and arrangement of individual cells play a role in shaping organ functions. However, analyzing multicellular features and exploring their connectomes in centimeter-scale plant organs remain challenging. Here, we established a set of frameworks named large-volume fully automated cell reconstruction (LVACR), enabling the exploration of 3D cytological features and cellular connectivity in plant tissues. Through benchmark testing, our framework demonstrated superior efficiency in cell segmentation and aggregation, successfully addressing the inherent challenges posed by light sheet fluorescence microscopy imaging. Using LVACR, we successfully established a cell atlas of different plant tissues. Cellular morphology analysis revealed differences of cell clusters and shapes in between different poplar (Populus simonii Carr. and Populus canadensis Moench.) seeds, whereas topological analysis revealed that they maintained conserved cellular connectivity. Furthermore, LVACR spatiotemporally demonstrated an initial burst of cell proliferation, accompanied by morphological transformations at an early stage in developing the shoot apical meristem of Pinus tabuliformis Carr. seedlings. During subsequent development, cell differentiation produced anisotropic features, thereby resulting in various cell shapes. Overall, our findings provided valuable insights into the precise spatial arrangement and cellular behavior of multicellular organisms, thus enhancing our understanding of the complex processes underlying plant growth and differentiation.

Large-volume fully automated cell reconstruction establishes a 3D cell atlas of poplar seeds, revealing the precise arrangement and topological organization of different cell types.

Introduction

Cellular geometric shapes play a substantial role in the biological process, serving as one of the critical means through which cells adapt to and respond to external changes (Arteaga et al. 2022). During development, the synergistic regulation of multiple gene functions plays a crucial role in promoting cell growth, differentiation, and adaptation to the environment, thereby leading to significant changes in cell shapes (Mathur et al. 2003). These changes contribute to the establishment of cell geometric anisotropy, which, in conjunction with the cell growth rate, ultimately determines the morphogenesis of tissues and organs (Erguvan et al. 2019). Moreover, multiple cell types with diverse geometric shapes are assembled into a biological system, resulting in an incredibly complex biological network. This cell network endows high-order functionality to maintain homeostasis for development and environmental adaptation (Arendt et al. 2016; Bassel 2019). In the field of neuroscience, establishing a multimodal database and cell atlas has provided a whole new paradigm for understanding the neuron phenotypic properties and activity of central neural networks (Milyaev et al. 2012; Bates et al. 2019). Likewise, in the study of the human brain, the Human Brain Project was established in several countries with the aim of providing a comprehensive analysis and understanding of the relationships between human brain phenotype, connection, function, psychological activity, and human behavior (Martin and Chun 2016; Poo et al. 2016). In plant science, a few reports have investigated the topological properties of the Arabidopsis (Arabidopsis thaliana) hypocotyls (Hys) and cellular growth dynamics in the Arabidopsis stamen based on cell atlas (Jackson et al. 2017; Silveira et al. 2022). Therefore, the integration of cell geometric shapes, gene expression levels, and cellular connectivity into a cell atlas can provide a comprehensive reference for examining the intricate variations and interplays among cells within the dynamic development and the morphogenesis mechanisms of biological systems (Zhang et al. 2021).

High-resolution 3D imaging is critical to visualizing and quantifying the internal cytological structure of organisms, thereby exploring the cell phenotypes and connectivity of complex multicellular sets (Zhang et al. 2019, 2021). In the past years, advances in several 3D imaging techniques have contributed to 3D reconstruction and multiscale network analysis in plant biology (Montenegro-Johnson et al. 2015; Ovečka et al. 2022a). At the microscale, electron microscopy and electron tomography (ET) can provide 3D ultrastructural data with unprecedented details for cells and organelles at subcellular resolution (Cui et al. 2019; Wang et al. 2020; Guo et al. 2023). However, the extreme expense, laboriousness of serial sectioning, and size limitation hinder their application for centimeter-scale plant organs (Shen et al. 2020). Similarly, laser scanning confocal microscope (LSCM) can be used for 3D imaging of Arabidopsis Hys, leaves, and pistils instead of larger samples due to its low scanning speed and small field of view (Kurihara et al. 2015; Ripoll et al. 2019; Ovečka et al. 2022b). Therefore, this requires a set of practical 3D imaging techniques for centimeter-scale plant organs. Furthermore, deep-learning algorithms for cell segmentation and morphometry are essential methods toward various cell identifications, utilizing a set of cytological descriptors to accurately describe and quantify cell phenotypes (Carpenter et al. 2006; Legland et al. 2016; Govek et al. 2023). In light of this foundation, single-cell morphology analysis based on high-resolution 3D imaging and deep-learning algorithms provides a robust analytical method and a comprehensive reference strategy for identifying, clustering, and analyzing a large number of cellular features in consecutive sections, enabling a thorough examination of diverse cell types and providing profound insights into their respective forms (Jo et al. 2021; Vergara et al. 2021).

Here, as part of a series of investigations on 3D imaging and quantitative analysis in plant organs, we build upon optimized imaging methods and deep-learning algorithms to accomplish 3D cellular imaging and accurate analysis of plant organs at a centimeter scale, such as different poplar seeds (Populus simonii Carr. and Populus canadensis Moench.) and shoot apical meristem (SAM) tissues of developmental Chinese pine seedlings (Pinus tabuliformis Carr.) (Shen et al. 2020; Zhang et al. 2020; Qi et al. 2022; Ma et al. 2024). Initially, we conducted the segmentation and identification of cell types from X-ray image stacks of poplar seeds. 3D cell visualization and feature analysis provided evidence for the feasibility of classification based on cell features. Building upon this foundation, we established a set of frameworks, large-volume fully automated cell reconstruction (LVACR), including light sheet fluorescence microscopy (LSFM) imaging, deep-learning algorithms, and cell feature analysis. Benchmark testing showed that LVACR exhibited high segmentation and aggregation efficiency. Using LVACR, we precisely segmented 10,000 of cells from LSFM image stacks and measured 80 cell morphological features for different poplar seeds and SAM tissues of different developmental stages in Chinese pine seedlings, which allowed us to establish comprehensive cell morphological atlases. Topological analysis in different seeds showed that radicle (Ra) cells had more adjacent cells and were more positioned along the shortest paths, while SAM cells exhibited opposite properties. We also observed that 2 individual cotyledons (Cos) exhibited identical feature distributions and connective patterns. Comparative analysis of different poplar seeds revealed the plasticity of cell shape and the robustness of cellular connections across different genetic backgrounds. Furthermore, an analysis of cellular shape dynamics in Chinese pine seedlings reveals that a rapid division of cells, also considered an initial proliferation burst, occurred within SAM cells of Chinese pine, triggering a morphogenetic transition. As these cells progress into the differentiation phase, the polar differentiation of SAM cells promoted the generation of a wider array of cell morphologies, thereby influencing the establishment of organ polarity. Our study provided a set of efficient frameworks for constructing cellular morphological atlases in plant organs at the scale of centimeters, facilitating the exploration of cell features.

Results

Feature-based cell type classification by 3D X-ray imaging, segmentation, and identification of P. simonii Carr. seeds

To verify the feasibility of classifying cells according to phenotypic characteristics, we explored the 3D cell structure in situ of the P. simonii Carr. seed by X-ray imaging, precise segmentation, and cell type identification. We initially used micro-computed tomography (micro-CT) imaging to rapidly assess the staining effect of several coloring agents on the P. simonii Carr. seed (Supplementary Fig. S1, A to D). Comparative analysis with the control group showed that La³⁺ exhibited poor penetration, with the seed coat showing the highest concentration, whereas Yb³⁺ staining had a limited effect (Fig. 1A). In contrast, we observed that the coimpregnation of CsI/FAA had the most significant coloring effect compared to other pretreatment methods, which was further confirmed by calculating and analyzing grayscale values (Fig. 1B). Following the CsI/FAA coimmersion protocol, a partial scan of the seed was performed using X-ray microscopy, resulting in the acquisition of over 2,000 continuous virtual slices with exceptional resolution (Supplementary Fig. S1D). Compared with semithin sections, various cells and organelles, such as Co cells, Hy cells, provascular cells, SAM cells, cell walls, and even intercellular spaces, could be distinguished from X-ray images (Fig. 1, C and D). We then cropped several tissues from X-ray image stacks and manually segmented more than 800 cells to precisely define cell types and analyze morphological features (Fig. 2, A and B; Video 1). Quantitative analysis showed that the volume and area of SAM cells and epidermal (Ep) cells were significantly smaller than other cell types, while the sphericity values were larger, indicating a more spherical shape. In contrast, cells in the Co and Hy exhibited a wide range of volume, surface area, and sphericity, indicating a diversity of cell features (Figs. 2, C and D, and 3; Supplementary Data Set 1). To compare differences between Co cells and Hy cells, we measured ellipticity prolate and ellipticity oblate between these cells. The results showed that Co cells had larger ellipticity oblate values, while Hy cells had smaller ellipticity oblate values, implying the geometric difference of the ellipsoid (Figs. 2D and 3; Supplementary Data Set 1). Moreover, unsupervised clustering of segmented cells based on 5 features yielded 8 clusters (Fig. 2E). Among these clusters, 6 classes contained separate cell types, like SAM cells, Hy cells, Co cells, and Ep cells, while 3 additional clusters were occupied by Co cells with other cell types (Fig. 2E). 3D models of segmented cells showed that these cells in each cluster have unique cell phenotypes and significant size differences (Fig. 3A). Further clustering of Co cells revealed 6 subclusters, showing obvious differences among them (Figs. 2F and 3B). Similar subclusters were found in the cropped Hy and epidermis tissues (Figs. 2, G and H, and 3B). In contrast, SAM cells were classified into 2 subclusters, indicating no significant differences among SAM cells (Figs. 2I and 3B). Taken together, we explored the cell features and classification patterns of major cell types, demonstrating that it was feasible to identify and classify cell types according to various features.

Figure 1.

Micro-CT and X-ray imaging of P. simonii Carr. seed pretreated with CsI/FAA coimpregnating. A) False color rendering of bottom, middle, and top cross-sectional images pretreated by control, LaCl₃/FAA, YbCl₃/FAA, and CsI/FAA coimpregnating. The color scale bar represents the range of grayscale values of the images. B) Line plot analysis of bottom, middle, and top cross-sectional images pretreated by control, LaCl₃/FAA, YbCl₃/FAA, and CsI/FAA coimpregnating. C and D) Comparison of semithin sections C) and X-ray images D) in the P. simonii Carr. seed. FAA, formalin-aceto-alcohol; Hy, hypocotyl; Co, cotyledon; Ep, epidermis; SAM, shoot apical meristem; Ra, radicle; Pvc, provascular cells; Sc, seed coat; Is, intercellular space. All scale bars are 100 μm.

Figure 2.

Segmentation, identification, and morphological analysis of cell types in the P. simonii Carr. seed. A) 3D profile and cropped tissues for analysis of P. simonii Carr. seed. Scale bar: 200 μm. B) The segmented cell models of ROIs (110 μm × 110 μm × 110 μm) cropped from A). n, the number of segmented cells. Scale bars: 40 μm. The color scale bar represents the range of cell volume. C) Five features of segmented cells from ROIs, including volume, surface area, sphericity, ellipticity oblate, and ellipticity prolate, were counted and displayed by Imaris 9.6 (see Supplementary Data Sets 1 and 5 for detailed feature values and descriptions). All scale bars are 40 μm. The color scale bar represents the range of different features. D) Violin plots of 5 features of segmented cells between 4 different tissues and the schematic models of the sphere and ellipsoid. Cell number (n) = 880. E) Clustering analysis of 5 features in all segmented cells yielded 8 clusters with distinct properties. F to I) Clustering analysis of 5 features in segmented cells from Co F), Hy G), Ep H), and SAM I). All data were standardized by minimum–maximum normalization. Color scale bars E to I) represent the range of normalized values.

Figure 3.

Hierarchical clustering and 3D models of the segmented cells. A) 3D models of the segmented cells, which are parallel to readers. The pie chart represents the proportion of each cluster in the overall population. B) Cluster analysis and cell models of segmented cells from different tissues, including Co, Ep, Hy, and SAM. Scale bar: 500 μm³.

A set of practically technical frameworks for centimeter-scale plant organs

To further explore the 3D shape of cell populations, we established a set of practically technical frameworks named LVACR for plant tissues at the scale of centimeters, including LSFM imaging, cell segmentation based on optimized deep-learning algorithms, and cell phenotypic analysis. We first cleared the tissue with NaOH–chloral hydrate and stained it with mPS-PI, as previously reported (Haseloff 2003; Truernit et al. 2008). The seeds retained some autofluorescence after FAA fixation and alcohol dehydration. After removing lipids and proteins, the samples turn out to be completely transparent and free of autofluorescence. By mPS-PI staining, the fluorescence with uniform intensity of samples could be observed by the fluoroscope (Fig. 4A). Then, the processed samples were scanned using LSFM, providing micron-level resolution, and more than 400 images were reconstructed by Imaris 9.6 (Fig. 4B).

Figure 4.

The workflow of LVACR. A) A workflow of poplar (P. simonii Carr.) seed clearing and staining. LM, light microscope; FM, fluorescence microscope. Scale bar: 500 μm. B) Raw data with micron-level resolution by LSFM, z axis: 2 μm, y and x axes: 0.572 μm. More than 400 continuous images were reconstructed by Imaris 9.6. C) The overview of the Blind2Unblind method. It consists of 2parts: the training phase and the inference phase. Images were denoised and enhanced by this framework (detailed steps can be seen in Materials and methods). D) Overview of the Swin-ResU-Net-Dilated architecture. The input to our model is a series LSFM images with single channels. The Swin-ResU-Net-Dilated creates nonoverlapping patches of the input data and uses a patch partition layer to create windows with a desired size for computing the self-attention. The encoded feature representations in the Swin transformer are fed to a Res-decoder via skip connection at multiple resolutions. Seg Head denotes the final segmentation layer. The final segmented output consists of 3 output channels, corresponding to affinity maps in the z, y, and x directions. E) 3D cell aggregation by Multicut, including oversegmentation of cell boundaries based on distance transformation using the 2D watershed algorithm, the probabilistic prediction of merging the corresponding 2 superpixels by training a RF classifier for each edge, and the MCMP solved by GAEC. Representing every cell with a random color. F) The strategy of block-wise segmentation and subsequent multiple block stitching.

For high-precision segmentation and identification, we integrated a set of optimized deep-learning algorithms combining the Blind2Unblind method for image denoising and enhancement, Swin-ResU-Net-Dilated for image segmentation, and Multicut for 3D aggregation. We first used Blind2Unblind as a lossless denoising method that can effectively address the issue of input loss while also dealing with pixel-independent and structured noise (Fig. 4C). This method included a global perception mask mapper based on mask-driven sampling and a nonblind lossless training strategy based on recoverable visibility loss. Specifically, we divided each noisy image into several blocks and set specific pixels in each block as blind spots. A global mask block was created as input, consisting of a sequence of images with masks (Supplementary Fig. S2). The block with the global mask was then sent into the network of the same batch to perform denoising, exploiting the physically independent noise context and the signal context-dependent properties. The global mapper sampled the denoised block at the blind spots and projected it onto the same plane to generate the denoised images (Fig. 4C; Supplementary Fig. S2).

Inspired by the prominent role that transformer-based models have demonstrated in various fields such as natural language processing and computer vision (Liu et al. 2021; Cardoso et al. 2022; Dufter et al. 2022; Hatamizadeh et al. 2022), we developed a U-shaped network (Swin-ResU-Net-Dilated) that combined Swin Transformer modules, dilated convolutions, and residual modules to achieve automatic cell wall segmentation. This network utilized 3D Swin Transformer modules as its encoding structure for downsampling, progressively revealing environmental information and expanding the receptive field. A ResNet based on dilated convolutions served as the decoding structure, gradually upsampling to restore image precision and communicating context information via skip connections at different resolutions. The architecture of Swin-ResU-Net-Dilated is illustrated in Fig. 4D. The Swin Transformer modules leveraged Window-based Multi-head Self-Attention (W-MSA) and Shifted Windows Multi-head Self-Attention (SW-MSA), achieving near-global attention capabilities while significantly reducing computational complexity. This transformation shifted from a square relationship with image size to a linear relationship, leading to a substantial decrease in the computational burden and an enhancement in the inference speed of models. Within each Swin Transformer block, a feature fusion method, referred to as Patch Merging, was employed to perform downsampling after each feature extraction. This expansion of the receptive field facilitates subsequent window attention operations on the original image, enabling multiscale feature extraction. This attribute made Swin Transformer modules particularly effective in tasks demanding dense prediction.

After obtaining the results of boundary segmentation, the cells have not been aggregated and instantiated yet. The objective of the aggregation operation is to assign a unique label to each independent cell, completing the reconstruction of the individual cell. To achieve this, we utilized the multicut solving algorithm for implementing 3D polymerization of the divided cells. Specifically, we first oversegmented the cell wall probability map based on distance transformation using the 2D watershed algorithm (Meyer 1994). Then, we predicted the probability of merging the corresponding 2 superpixels by training a random forest (RF) classifier for each edge (Beier et al. 2017). Finally, we modeled the problem of superpixel aggregation as the minimum cost multicut problem (MCMP) and solved it using the greedy additive edge contraction (GAEC) algorithm (Supplementary Table S1) based on edges (Keuper et al. 2015; Beier et al. 2017), ultimately obtaining the 3D instance structure of the cell (Fig. 4E). Limited by computing resources, it is impossible to segment large amounts of data directly. Therefore, we adopted a typical reconstruction pipeline, as shown in Fig. 4F, which involved dividing images into multiple blocks and independently 3D reconstructing them (in-block segmentation). Then the segmented adjacent blocks were merged so that the corresponding cells were stitched. Finally, tens of thousands of cells were segmented from LSFM data.

Evaluation of the denoising and segmentation performance of LVACR based on benchmark testing

To verify the denoising performance of Blind2Unblind, we conducted comparative tests across different denoising methods on public data sets such as CREMI (Funke et al. 2016) and SNEMI3D (Kasthuri et al. 2015). Comparing to the traditional unsupervised method BM3D (Dabov et al. 2007) and 4 self-supervised denoising approaches: Noise2Void (N2V) (Krull et al. 2019), Laine19-mu, Laine19-pme (Laine et al. 2019), and NBR2NBR (Huang et al. 2021), simulated denoising results showed that Blind2Unblind delivered optimal values in peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) under various conditions, such as static Gaussian noise (σ = 25, 38), dynamic Gaussian noise (σ ∈ [5, 50]), static Poisson noise (λ = 10, 30), and dynamic Poisson noise (λ ∈ [5, 50]) (Supplementary Tables S2 and S3 and Figs. S3 and S4). Additionally, compared to supervised baselines N2C (Ronneberger et al. 2015) and N2N (Lehtinen et al. 2018), Blind2Unblind also provided comparable or superior values in PSNR and SSIM, demonstrating the exceptional performance of the lossless self-supervised denoising framework (Supplementary Tables S2 and S3 and Figs. S3 and S4). Visualization results on real Arabidopsis images revealed that despite the low noise levels present in the actual images, the Blind2Unblind method still offered higher recovery precision than other self-supervised denoising methods, especially showing a remarkable advantage in restoring the fine and high-quality structures (Supplementary Fig. S5).

To further showcase the robustness and effectiveness of the Swin-ResU network, we also conducted comparative experiments on the 2 public data sets: the Medical Segmentation Decathlon (MSD) data set for Hepatic Vessels (Antonelli et al. 2022) and the BraTS21 data set (Baid et al. 2021). Comparing our method with other methods, we found that the average Dice scores of Swin-ResU-Cnn on the MSD data set outperform nnUNet (Isensee et al. 2021), nnFormer (Liu et al. 2022), and TransVW (Haghighi et al. 2021) by 0.06%, 3.90%, and 2.64% for liver segmentation and 14.07%, 0.13%, and 7.16% for liver tumor segmentation, respectively (Supplementary Table S4). Benchmark tests on the BraTS21 data set showed the Dice scores for all tumor regions of both Swin-ResU-Cnn and Swin-ResU-Trans outperformed competing methodologies. Specifically, Swin-ResU-Cnn outperformed nnUNet, nnFormer, and TransBTS (Liu et al. 2018) by 0.22%, 2.22%, and 2.54%, respectively (Supplementary Table S5). These results further validated the effectiveness and generalizability of our proposed architecture across different tasks and imaging modalities. In addition, Hausdorff distance (HD) results indicated that in most liver and brain tumor segmentation tasks, Swin-ResU-Cnn achieved higher HD scores than Swin-ResU-Trans. In terms of parameters, Swin-ResU-Cnn and Swin-ResU-Trans models were of moderate size compared to larger models like nnFormer. Furthermore, Swin-ResU-Trans had the smallest GFLOP compared to other methods, enabling higher computational efficiency and acceptable accuracy (Supplementary Tables S4 and S5). These results further verified the robustness and effectiveness of the Swin-ResU network.

Furthermore, we conducted comprehensive tests on the denoising, segmentation, and aggregation performance of LVACR to evaluate the high efficacy and robustness of this framework in handling the LSFM data set. We first evaluated the denoising performance of the Blind2Unblind method. From Fig. 5A, we observed that LSFM images enhanced by Blind2Unblind exhibited higher clarity and more pronounced contrast in gray values compared to others, suggesting the outstanding denoising performance of Blind2Unblind. We then evaluated the effect of several denoising methods on cell segmentation and aggregation by employing common evaluation metrics such as Intersection over Union (IOU, including precision, recall, and accuracy), Variation of Information (Vinfo) (Meilă 2007), and Adjust Rand Error (ARAND) (Arganda-Carreras et al. 2015). The results showed that image segmentation and aggregation based on Blind2Unblind denoising exhibited superior performance, especially in areas of high signal within images, with higher precision, recall, and accuracy, as well as lower Vinfo and ARAND values (Supplementary Fig. S6, A and B, and Data Set 2). After that, we further evaluated the segmenting and aggregating performance of Swin-ResU + Multicut. Compared to the commonly used CNN network for segmentation, we observed that Swin-ResU contributed to reducing segmentation defects and omissions, although there were still a few incorrect segmentations (Supplementary Fig. 5B and S7A). Compared to Waterz and Mutex + Watershed for aggregation, the Multicut solving algorithm generated fewer aggregation errors (Fig. 5C; Supplementary Fig. S7B). The results of general evaluation metrics indicated that the integrated framework of Blind2Unblind + Swin-ResU + Multicut had good segmentation performance, with an average precision of 0.9835, an average recall of 0.9073, and an average accuracy of 0.912. The lower Vinfo and ARAND values indicated fewer false splits and merges (Fig. 5D; Supplementary Data Set 3). These findings provided further evidence of the significant performance of LVACR in handling the LSFM data.

Figure 5.

Enhancement effect of different LSFM images and comparison of effects for cell segmentation and aggregation based on different algorithms. A) Comparison of different slices by no denoising, Gaussian denoising, and Blind2Unblind denoising. Line plots represent the statistics of gray values. Scale bar: 100 μm (top) and 25 μm (bottom). B and C) Cell segmentation and aggregation by different methods, like CNN + Waterz (0.10), CNN + Multicut, CNN + Mutex + Watershed, Swin-ResU + Waterz (0.10), Swin-ResU + Mutex + Watershed, and Swin-ResU + Multicut (ours). The arrows in B) indicate segmentation errors. The arrows in C) indicate aggregation errors. No cell walls were displayed in C) for a clearer representation. Representing every cell with a random color. Scale bars: 100 μm. D) Quantitatively evaluate the efficiency of these methods. Boxplots represent mean, 25th, and 75th quartiles; whiskers represent minimum and maximum (n = 5). Detailed evaluation metrics include precision, accuracy, recall, Vinfo, and ARAND (see Materials and methods for detailed descriptions of these metrics).

More importantly, we compared our integrated approach with existing methods for plant cell reconstruction in the field, including TWANG (Stegmaier et al. 2014), Supervoxel-CNN (Stegmaier et al. 2018), Cellpose (Stringer et al. 2021), and PlantSeg (Wolny et al. 2020). The comparison results of the truth values indicated that our approach consistently outperformed the above methods in terms of accuracy by an average of 13.48%, 3.56%, 4.92%, and 1.43%, respectively. Regarding precision, our approach achieved higher values than the methods mentioned above by 13.46%, 9.42%, 2.6%, and 2.85%, respectively (Supplementary Fig. S8, A and C, and Data Set 4). When it came to recall, our methods exhibited a higher value that was only behind Cellpose, which exhibited significant identification errors (Supplementary Fig. S8). Moreover, our methods also exhibited the lowest values in terms of Vinfo and ARAND metrics, which indicated the least amount of splitting and merging errors yielded (Supplementary Fig. S8, B and C, and Data Set 4).

Cell identification from clustering features of a 3D morphological atlas

Following the high-efficiency segmentation and identification of LSFM image stacks, 3D models containing morphological data of segmented cells were created (Fig. 6A; Video 2). After that, we planned to construct a cell phenotype atlas according to various features. Initially, we quantified 80 feature descriptors of segmented cells for P. canadensis Moench. and P. simonii Carr. seeds and subsequently constructed separate 3D shape atlas for each (Fig. 6B; Supplementary Data Sets 5 and 6). Additionally, through unsupervised clustering using Seurat (Hao et al. 2021) and plotting via uniform manifold approximation and projection (UMAP) (Becht et al. 2019), 9 distinct clusters were identified in both P. canadensis Moench. and P. simonii Carr. seeds (Fig. 6B). Upon creating the cell atlas for P. simonii Carr., we found that these clusters showed diverse identities. For instance, cells classified in clusters 0, 4, and 5 displayed smaller surface area, volume, equivalent diameter (eqdiameter), and larger elongation values compared to others, while cells in cluster 7 exhibited contrasting characteristics (Supplementary Fig. S9, A and C, and Data Sets 5 and 6). Similar results were also observed in the P. canadensis Moench. cell atlas, such as smaller volume, area, and eqdiameter for cells in cluster 7; larger volume and area for cells in clusters 4 and 8; and smaller elongation for cells in clusters 3 and 8 (Supplementary Fig. S9, B and C, and Data Sets 5 and 6). These results demonstrated that conducting clustering analysis based on diverse shape features can assist in identifying a wider range of cell shapes and quantifying the differences among them.

Figure 6.

Cell shape clustering and topological analysis of segmented cells in different poplar seeds. A) 3D cell model of segmented cells in P. simonii Carr. and P. canadensis Moench. seeds. Representing every cell with a random color. Scale bars: 100 μm. B) UMAP plot of segmented and filtered cells in P. simonii Carr. and P. canadensis Moench. seed based on features (see Supplementary Data Sets 5 and 6 for detailed feature descriptions and values), colored by clusters (c0 to c8). C and D) The defined cell types were annotated and colored on the UMAP plot in P. simonii Carr. and P. canadensis Moench. seed. E) The relative frequency of cell degree and BC values among different cell types in P. simonii Carr. and P. canadensis Moench. seeds (see Supplementary Data Set 7 for detailed values). Cell numbers (n) = 28,034, 27,693, 55,653, and 54,977, respectively. F) 2D rendered images with cell area, ellipticity oblate, and local density (Voronoi diagram) of Cos in poplar seeds. Scale bars: 100 μm. Color bars represent the range of different features. G) Relative frequency of cell degree and BC values of 2 individual Cos in different seeds. The data were examined by the Mann–Whitney–Wilcoxon test. ns, no significant difference.

Integration analysis of cell types and features underlying connective properties in different seeds

Based on cell spatial distribution, we annotated major cell types in the cell atlas to verify the feasibility and identity of the clusters in different poplar seeds. From the annotated atlas and statistical results of the P. simonii Carr. seed, we found that cells in the Co and Hy tissues were distributed in all clusters, consistent with the distribution of these cells in the P. canadensis Moench. atlas, showing significant differences in cell shapes and overlapping distribution patterns in these tissues (Fig. 6, C and D; Supplementary Fig. S10 and Data Set 6). Likewise, we also found SAM, Col, and Ep cells were mainly classified into the same clusters (C0, C3, and C4 in the P. simonii Carr. atlas; C0, C2, C5, and C7 in the P. canadensis Moench. atlas) in both atlases, indicating a similar cell shape classification pattern (Fig. 6, C and D; Supplementary Fig. S10 and Data Set 6). In contrast, we found that Ra cells exhibited a predominant clustering pattern in the P. simonii Carr. (cluster 7) and P. canadensis Moench. (cluster 1) atlas, indicating relatively consistent cell shape among these cells with large volume, area, and small elongation values (Supplementary Figs. S9C and S10 and Data Set 6).

We next rendered several images to visualize the morphological differences between the P. simonii Carr. and P. canadensis Moench. seeds at the cellular level. The results showed that the 2D cell area of Ep cells, Co cells, Hy cells, and Ra cells in P. simonii Carr. was generally larger than that of P. canadensis Moench., whereas there were no significant differences in SAM cells or Col cells (Supplementary Fig. S11A). To precisely quantify differences between both seeds, we compared major 3D features of segmented cells and the results showed that volume, 3D area, eqdiameter, perimeter, and area fraction of Co, Hy, and Ra cells in P. simonii Carr. were larger than those in P. canadensis Moench. (Supplementary Fig. S11, B to F). Statistical analysis of all P values of 80 features for cells between these 2 species showed that there were no significant differences in Col and SAM cells between both seeds (most P ≥ 0.05). Conversely, there are significant differences in Hy, Co, Ep, and Ra cells between the 2 seeds (most P < 0.001), demonstrating the shape plasticity of these cells within seeds between different genetic backgrounds (Supplementary Fig. S11G).

Using the network quantitative analysis strategy as previously reported (Jackson et al. 2017), we calculated and compared the degree and betweenness centrality (BC) values of segmented cells to explore the topological properties of these cells in different tissues. The statistical results in the P. simonii Carr. cell network showed that the degree of cells in SAM was smaller (most degree values < 15) and the degree of cells in the Ra was greater than other cell types (most degree values ≥ 15) (Fig. 6E; Supplementary Data Set 7). Similarly, it has been observed that SAM cells have lower BC values (over 80% BC values < 2 × 10⁻⁴) and Ra cells have larger BC values. Lower degree and BC values are also observed in Ep cells and Col cells (Fig. 6E; Supplementary Data Set 7). Moreover, similar distribution patterns of BC and degree in different tissues were also present in the P. canadensis Moench., indicating a conservative pattern of cellular connectivity between poplar seeds across different genetic backgrounds (Fig. 6E; Supplementary Data Set 7).

The homogeneous Cos formed by a high similarity of geometrical morphology of tissues, cell shape, and connectivity

Morphology symmetry exists commonly in organisms with bilateral structures. To assess whether the two Cos of bilateral plants are consistent, we reconstructed 3D structure models and analyzed the geometrical morphology (3 biological replicates) of the 2 Cos in the different seeds. The results showed that there are no significant differences in volume, area, sphericity, ellipticity oblate, or ellipticity prolate between the 2 Cos (Supplementary Fig. S12, A and B, and Data Set 8). From the 2D rendered images with cell area, ellipticity oblate, and local density in Cos, we found that there are symmetrical distribution patterns of cells across 2 individual Cos (Fig. 6F). To further determine if there exist differences between 2 individual Cos at the cellular scale, we compared the cell features and distribution in the cell atlas between Co-1 and Co-2. The statistical results of P. simonii Carr. cells showed that cells in 2 Cos exhibited a consistent distribution of major features like area3d, elongation, volume3d, and so on. Although there were slight proportion differences observed in different clusters, ranging from 42.9% to 56.2% and 43.8% to 57.1% in proportion of Co-1 and Co-2 cells, respectively, these differences were found to be not statistically significant (Supplementary Fig. S12C). We also investigated the topological properties of cells between 2 Cos in P. simonii Carr. seeds. The results showed that the distribution patterns were similar in degree and BC values between Co-1 and Co-2 cells, which were consistent with local cell density (Fig. 6G; Supplementary Data Set 7). In addition, similar results were also found in the P. canadensis Moench. Co tissues (Fig. 6G; Supplementary Data Set 7). These results demonstrated a homogeneous pattern between 2 individual Cos in terms of the geometric morphology of tissues, cellular shapes, and connectivity.

A spatiotemporally resolved dynamic atlas of developing SAM in the seedlings of Chinese pine by LVACR

We initially examined the developmental stages of SAM tissues in situ. This observation encompassed the initiation of leaf primordia (stage 1, 8 days after germination, DAG 8 for short), the development of leaf primordia (DAG 9, 10, and 12), and the morphological establishment of young leaves (DAG 15) (Fig. 7, A and B). Quantitative analysis revealed a significant increase in the volume of SAM tissues during stages 3 to 5 (Fig. 7B). After that, we segmented and identified cells in the SAM tissue of Chinese pine seedlings. The results showed that LSFM images processed by the LVACR framework exhibited a significant decrease in image noise and a significant enhancement in contrast compared to the raw images, with particularly notable noise reduction effects in areas of dense signaling (Supplementary Fig. S13, A and B). We also found high accuracy and precision in cell segmentation and identification, demonstrating the high efficiency and robustness of extracting and analyzing cellular morphology from large-scale data sets (Supplementary Fig. S13, C and D). Subsequently, we conducted a detailed quantification and evaluation of the dynamic changes in multicellular patterns of SAM tissues during the cell division and differentiation processes using LVACR framework (Fig. 7, C and D). Based on this analysis, we generated 3D thermal maps portraying cellular traits, leveraging spatial coordinates of segmented cells and normalized feature values (Fig. 7E). The results showed a drastic reduction of cell volume3d and area3d, coupled with a marked increment in cell count during stage 3. This was followed by a rapid expansion in both cellular volume3d and area3d, along with a pronounced increase in cell numbers during stages 4 and 5 (Fig. 7, D and E; Supplementary Fig. S14 and Data Set 9). Regarding elongation values, where lower values indicated more elongated cells, a significant peak was observed at stage 3, which then gradually decreased in stages 4 and 5, showing an inverse relationship with cell anisotropy values (Fig. 7, D and E). Collectively, these results suggested that an initial burst of cell proliferation during stage 3 led to a substantial increase in cell counts while influencing cell shapes, subsequently transitioning toward elongation and anisotropy in stages 4 to 5. Additionally, we observed that cells in the tip region of the young leaves showed a distinct trend toward pronounced anisotropy and elongation compared to basal cells during stage 5. Furthermore, cells at the lateral tip region exhibited augmented volume3d, area3d, and anisotropy, along with lower elongation values, indicating a more elongated shape. These results suggested that polarity establishment is related to significant morphological differences among cells in different regions (Fig. 7E; Supplementary Fig. S15).

Figure 7.

Development patterns of the SAM of young Chinese pine seedlings. A) SAM tissues of young Chinese pine seedlings at dynamical developmental stages (DAG 8, 9, 10, 12, and 15). White arrows: leaf primordial. Scale bars: 1 mm (left) and 100 μm (middle and right). B) 3D visualization and morphological changes of SAM tissues at different developmental stages (S1, S2, S3, S4, and S5). Data are means ± Sd (n = 2). Scale bar: 100 μm. C) 3D cell models of SAM tissues at different stages. Scale bar: 100 μm. Representing every cell with a random color. D) Statistical analysis of cell features (including volume3d, area3d, elongation, and anisotropy) of segmented cells. Elongated cells have small values close to 0. Data are means ± Sd (cell numbers [n] = 482, 586, 981, 2,580, and 10,591, respectively). ns, no significant; asterisks denote statistically significant differences (**P < 0.01, ***P < 0.001, ****P < 0.0001); the data were examined by Mann–Whitney–Wilcoxon test. E) The 3D scatter heat maps are colored by the normalized values of cell features. The scatters represent the barycenter coordinates of cells. The histograms depict a dynamic shift of cell features. Cell numbers in different stages (n) = 482, 586, 981, 2,580, and 10,591, respectively. DAG, days after germination.

By using a set of 80 cell features based on LVACR, we constructed a cell shape atlas of SAM cells across 5 stages of SAM development, clustering these cells into 15 distinct clusters (Fig. 8A; Supplementary Data Set 9). It was observed that cells during stages 1 to 3 were predominantly classified as clusters 1, 2, and 9. As development progressed, the diversity of clusters augmented, culminating in 10 clusters by stage 5, which facilitated the emergence of a more diverse range of cell shapes (Fig. 8, A to C). We also analyzed the local density of SAM cells at different stages, and the results showed that the cell density at stage 3 was higher compared to the first 2 stages, suggesting a surge in cell division at this juncture. Subsequently, regions of low density began to proliferate during stages 4 and 5, implying that the cells within these regions commenced their growth and differentiation (Fig. 8D). In summary, our results showed a cellular growth model for SAM cells in seedlings (Fig. 8E). During the development of SAM, these cells transitioned into an initial “proliferation burst” stage (stage 3) and underwent shape transformation in the early stage, characterized by smaller cell sizes, reduced anisotropy, and a tendency toward nonelongation. Over the following 2 stages, SAM cells gradually differentiated and developed anisotropic features, leading to the emergence of a greater diversity of cell shapes.

Figure 8.

Cell clustering and dynamic development models of SAM cells at different stages. A) Clustering of segmented cells in different stages based on 80 features. B and C) The histogram depicts a dynamic shift in the proportion of cells across various clusters at different developmental stages. The line chart showed that the number of clusters increased gradually (the clusters that exceeded the 2% threshold are considered effective). Cell numbers in different stages (n) = 482, 586, 981, 2,580, and 10591, respectively. D) The graphical representation shows localized cellular density at different stages, with high-density regions indicating actively proliferating cells and low-density areas corresponding to cells primarily in growth and differentiation phases. Scale bar: 30 μm. E) A schematic representation of the multicellular growth model of SAM.

Discussion

Advancements in 3D imaging have enabled us to observe plant internal structures in unprecedented ways (Bassel 2015). This progress echoed the need for a botanical renaissance proposed by Ledford (2018), particularly in the 3D analysis of plant morphology and anatomy. However, due to the lack of effective pretreatment for complex tissues, 3D imaging of centimeter-scale plant organs at cell resolution still remained a challenge. In animals, the sample processing strategies of contrast enhancement and tissue optical clearing provide insights into the 3D structure and development of the whole animal embryo (Brasch et al. 1984; Zhu et al. 2020; Gong et al. 2023). In plant biology, multiview imaging obtained by micro-CT or X-ray microscopy was effectively applied for various organs at the cellular or tissue level (Wang et al. 2017; Qi et al. 2022; Trueba et al. 2022). Additionally, translucent model plant tissues or organs can be examined by means of LSCM and LSFM with cellular resolution (Truernit et al. 2008; Maizel et al. 2011; Janes et al. 2018). In this study, we optimized an effective pretreatment method for X-ray scanning. Compared with other pretreatment methods mentioned above, we found imaging resolution has been considerably improved using CsI/FAA coimmersion and critical point drying, as confirmed by semithin sections. Furthermore, after evaluating the effect of NaOH–chloral hydrate clearing and mPS-PI staining, we validated that this protocol was rapid and efficient for high 3D imaging and precise cell recognition of plant organs at the scale of centimeters.

Advancements in 3D imaging technology have generated unprecedentedly complex and large-volume data sets, making it challenging to achieve accurate segmentation and cell identification (Magness et al. 2024). Some advanced deep-learning algorithms have been developed for 3D reconstruction and analysis of high-throughput data sets in biomedicine (Scheffer et al. 2020; Jiang et al. 2021; Jo et al. 2021; Govek et al. 2023). However, a critical limitation remained in their efficient application to LSFM data. Different from other imaging techniques, LSFM imaging faced unique challenges that impeded accurate identification of cellular biological structures, including pixel-independent noise, significant distortion and drift when capturing repeated shots at fixed positions, and a much higher resolution along the x–y axis compared to the z axis in serial images. These issues collectively hindered the accurate identification of cellular biological structures. In our study, we first applied the Blind2Unblind method to address the inherent issues of structured noise and low signal-to-noise ratio (SNR) in LSFM imaging. Benchmark testing has shown that Blind2Unblind is more suited for data with complex and variable noise patterns, successfully overcoming the problem of input loss in existing self-supervised denoising methods. In addition, we also propose a comprehensive approach that combines cell segmentation and aggregation, namely Swin-ResU + Multicut. In comparison to other methods evaluated in benchmark tests, Swin-ResU + Multicut exhibited excellent performance in both segmentation and aggregation, effectively minimizing errors in these processes. Although the common segmentation model, Cellpose, can bypass the training step and deliver satisfactory results for diverse data, the training-based approach of Swin-ResU + Multicut offers enhanced efficiency and reliability in segmentation and aggregation, particularly for anisotropic data. Moreover, the strategy of block-wise segmentation and subsequent multiple block stitching effectively tackled the challenge of segmenting large-scale and voluminous data sets. Overall, specific calibration for large-volume data sets and targeted optimization for complex noise patterns successfully addressed the inherent challenges posed by LSFM imaging, indicating our methods were particularly suitable for LSFM data sets.

Various cell types in biological systems have complex and diverse features that can be used to distinguish their inherent characteristics without additional information. In animals, a previous study showed that a cellular atlas was constructed and major cell classes in the Platynereis dumerilii were identified based on complex features of cells and nuclei (Vergara et al. 2021). In plant biology, previous studies in Hys and SAM tissues of Arabidopsis had shown that different cell types have distinct shapes and circumferential lengths, highlighting their specificity and wide spatial distribution (Kierzkowski et al. 2012; Montenegro-Johnson et al. 2015; Jackson et al. 2019). In this study, we first precisely quantified and identified major cell types based on X-ray data. Clustering and statistical analysis showed that SAM cells are smaller and show cytological homogeneity, in contrast to the larger and wider cell phenotype range of Co and Hy cells. These findings indicated the feasibility of cell classification based on diverse cytological features. Noticeably, when constructing cell phenotype atlases from LSFM data and analyzing the characteristics of clusters, we found more diverse features and types of Co cells and Hy cells, similar to the results obtained with the X-ray data analysis mentioned above. Our findings provided additional supporting evidence from a geometric morphometric perspective for the notion that Co and Hy cells possess a greater diversity of cell types. We also found that Ra cells mainly belonged to a unique cell cluster with features of large volume, area, eqdiameter, and elongation, revealing the unique elongation pattern of these cells, suggesting their potential role in nutrition absorption and transport. Furthermore, our study revealed that cell size varied significantly between different poplar seed species, indicating plasticity in cell shapes.

Plant complex architecture refers to the collection of the geometric and topological structures of various cells and tissues (Conn et al. 2017). Understanding topological properties is crucial to deciphering complex cellular patterns and multicellular structures (Duran-Nebreda and Bassel 2017; Zhang et al. 2021). Previous research demonstrated the robustness and adaptability of atrichoblast patterning across different genetic backgrounds using topological analysis (Jackson et al. 2017). In our study, we found that there were more neighboring cells in Ra cells. And these cells lie upon significantly shortest paths, which refer to the minimum number of distinct edges that pass through from one vertex to another in the network (Barthélemy 2011). These findings suggested that Ra cells exerted a crucial role in the cell network. Additionally, topological analysis revealed that there are similar geometric shapes, cellular feature distributions, and connectivity patterns within 2 individual Cos. Our findings provide geometric morphometric and topological evidence to support the idea that 2 individual Cos in the bilateral plants exhibit a highly symmetrical and uniform pattern. Moreover, we also observed similar cellular connectivity patterns between different seeds, despite exhibiting apparent cell morphological differences. These findings further demonstrated the robustness and adaptability of cellular connectivity across different genetic backgrounds and suggested the high efficiency of LVACR in processing LSFM data.

Different developmental patterns in plant cells can give rise to comparable geometrical shapes (Erguvan et al. 2019). This characteristic, along with an early stage “proliferation burst” and intercellular interactions, can swiftly drive the establishment of organ polarity (Bowman and Floyd 2008; Li et al. 2024). In this study, our findings revealed that an initial burst of cell proliferation occurs within SAM cells in stage 3, evidenced by a significant increase in cell number and a noticeable decrease in cell volume, leading to a nonpolarity shape. During the subsequent morphogenesis of young leaves, cells progressively increase in volume and surface area, adopting shapes that become more anisotropic. Notably, cells at the lateral tip not only increased in volume and surface area but also exhibited heightened anisotropy and elongation. The diverse cell shapes indicated a collective shift toward anisotropic differentiation, shaping the morphogenesis of young leaves. These observations suggested that cell differentiation drives cells to acquire specific shapes, leading to the development of a wider range of cell types and the rapid growth and establishment of polar shapes in young leaves.

Gene expression variations in multicellular organisms have a profound impact on cell development and organ morphogenesis (Arendt et al. 2016). A significant manifestation is that genes responsible for cytoskeletal organization regulators collectively govern the assembly and structure of the cell cytoskeleton, thus exerting precise control over the distinct morphology of cells (Mathur et al. 2003; Xu and Bassel 2020). Such regulatory mechanisms play a key role in shaping the intricate cellular landscapes of multicellular organisms. The advent of single-cell RNA sequencing (scRNA-seq) reshaped the scale of studying plant gene function, transitioning from the bulk level to the level of individual cells. In addition to the identification and differentiation of various cell types, scRNA-seq enabled us to uncover molecular mechanisms at the cellular level (Kim et al. 2021; Lee et al. 2023; Liang et al. 2023). More importantly, using single-nucleus RNA sequencing (snRNA-seq) to comprehensively survey the cell types and states during 10 developmental stages, from Arabidopsis seeds to siliques, generated important resources for the study of cell type regulation throughout the developmental process (Lee et al. 2023). However, due to the lack of 3D cytological characteristic data, accurately classifying cell types remains complicated (Morris 2019). Spatial transcriptomics provides in situ information such as cellular characteristics and spatial coordinates that cannot be obtained solely by scRNA-seq. Integration of these 2 techniques enabled us to gain insights into various aspects, including dynamic development mechanisms, and symbiotic processes (Liu et al., 2023). Indeed, spatial transcriptomics can only provide limited information due to its reliance on tissue sectioning (Cox et al. 2022). Therefore, a comprehensive data set of cell morphology is still required. On this basis, scRNA-seq can be utilized to conduct a comprehensive analysis of transcriptional changes at the single-cell level throughout development. High-resolution 3D imaging and segmentation algorithms can be employed to fully reveal the patterns of dynamic changes in individual cells. Furthermore, the transcriptional expression matrix and cellular features in spatial transcriptomics data can be used as a bridge to map and visualize the scRNA data and single-cell morphological data in 3 dimensions (Cox et al. 2022). This integration based on computational strategies enables us to unbiasedly define cell types, examine these genes involved in cytoskeletal formation, and track their expression level changes across different cell types during development, thereby uncovering the fundamental principles of how these genes regulate changes in 3D cytological features.

In conclusion, our study presented a 1-stop framework that integrated clearing and LSFM imaging of organs at the centimeter scale, automated segmentation, and 3D geometric analysis. Utilizing this comprehensive framework, we successfully identified 3D geometric features of multicellularity and provided a dynamic 3D view of cellular behavior in complex biological systems. These findings enhanced the understanding of complex cellular architectures in plants, serving as a significant achievement and valuable reference for exploring complex processes of plant growth and differentiation.

Materials and methods

Sample preparation, CT imaging, and 3D reconstruction

The detailed operation process for micro-CT scanning is as follows. The mature seeds were first soaked in FAA at 4 °C for 1 d and then transferred to a mixed solution of 10% (w/v) cesium iodide solution and FAA at 4 °C for 7 d. The samples were then sequentially transferred to a series of increasing ethanol concentrations (70%, 80%, 90%, 95%, 3 × 100% v/v) for dehydration, with each step lasting 30 min. To further eliminate the interference of moisture and free cesium on scanning imaging, the samples can be dried about 4 h by the CO₂ Critical Point Drier (Leica EM CPD300, Germany). After that, the samples were mounted in a plastic container and scanned by micro-CT (Skyscan 1172, Bruker, Belgium) with a resolution of 0.74 μm or X-ray microscope (Xradia 620 Versa, Carl Zeiss, Germany) with a resolution of 0.57 μm to obtain large amounts of 2D images. Finally, micro-CT images were imported into Imaris 9.6 (Bitplane AG, Switzerland) to reconstruct a 3D model of whole embryos for further analysis. To compare and quantify the differences among these contrast-enhanced scans, we used the open-source analysis software ImageJ (https://imagej.nih.gov/ij/download.html) (Schindelin et al. 2012) to perform pseudo-color rendering on selected regions and calculate gray values.

Cell segmentation, annotation, and analysis of cell types in P. simonii Carr. embryo

Several regions of interest (ROIs) of Hys, Cos, and SAM from X-ray images were cropped and quantitatively analyzed by Imaris 9.6. Firstly, import X-ray images into Imaris 9.6 and choose “Surface-create-Algorithm Settings” to crop ROIs. Then, use “Surface-draw” to segment and define 4 cell types, such as Ep cells, Hy cells, Co cells, and SAM cells. After that, count and output the parameters of these cells. Notably, manual segmentation was utilized in this context exclusively to verify the feasibility of feature-based cell classification. To ensure the accuracy of statistical data, it is necessary to eliminate incomplete cells in contact with the ROI border. Finally, the open-source analysis software GraphPad Prism version 9.0.0 (https://www.graphpad.com/features) was used to calculate and analyze these parameters.

Sample preparation, LSFM scanning, and 3D reconstruction

First of all, Chinese pine seeds were germinated in sphagnum moss media at 22 °C with a 16-h light/8-h dark cycle. After several days of germination, SAM tissues from seedlings at different stages were selected under a stereomicroscope. Next, poplar seeds (P. simonii Carr. and P. canadensis Moench.) and SAM tissues of Chinese pine seedlings (P. tabuliformis Carr.) were fixed in FAA fixative at 4 °C for at least 1 d. Subsequently, the samples were sequentially dehydrated step by step and were then immersed in petroleum ether at room temperature for 1 d to extract their lipids. After that, the samples were soaked in acetone and chloroform at room temperature for 12 h in proper order. To decolorize the samples and remove proteins, the rehydrating samples were soaked in 4% (w/v) NaOH at 60 °C for 2 d with solutions changed every 12 h until they were almost semitransparent. The semitransparent samples were then transferred to saturated chloral hydrate for 12 h at 60 °C to become completely transparent. After that, the samples were dyed with propidium iodide according to previously described methods (Truernit et al. 2008). The transparent seed embryos were hydroformed with 1 mL of periodic acid solution for 1 h and then rinsed with distilled water 3 times for 1 min each time. After adding 1 mL NaHSO₃-HCl solution and 1 mL PI solution for costaining for 3 h, the staining solution was discarded. The samples were rinsed with distilled water for 1 h and imaged at 488 nm laser by light sheet Z.1 from Carl Zeiss Microscopy. Finally, the image stacks and 3D model were generated by Imaris 9.6.

Image denoising and enhancing

This Blind2Unblind method was improved and performed based on the previous report (Wang et al. 2022). First, the global masker ${\mathbf{\Omega }}_{(\cdot )}$ adds blind spots to the noisy image $\mathbf{y}$ to create masked blocks. Next, the global perception mask mapper samples from the blind spots of the denoised blocks to get$h(f_{\theta }({\mathbf{\Omega }}_{\mathit{y}}))$. At the same time, the denoiser $f_{\theta }(\cdot )$ takes $\mathbf{y}$ as input and outputs the denoised result$f_{\theta }(\mathbf{y})$. The recoverable visibility loss uses the invisible term $h(f_{\theta }({\mathbf{\Omega }}_{\mathit{y}}))$ as a medium for gradient updates, achieving a transition from blind to visible. Additionally, a regularization term is used to stabilize training. After that, the trained denoising model was used for inference. The denoising network directly generates a denoised image from the noisy image $\mathbf{y}$, without the need for additional operations.

Here, we introduce how to use the Blind2Unblind approach for self-supervised denoising. As blind spot denoising acts as a medium, the training process should follow a transition from blind spots to nonblind spots. However, a naive recoverable visibility loss that only uses a single variable and optimizes both the blind spot term and the lossless visibility term through backpropagation can lead to unstable training. Therefore, a regularization term is introduced to constrain the lossy blind term and stabilize the training process. The recoverable visibility loss with regularization is as follows:

$$\begin{array}{c}\begin{array}{cc}\mathcal{L}& ={\mathcal{L}}_{\mathrm{rev}}+\eta \cdot {\mathcal{L}}_{\mathrm{reg}}\\ & =\Vert h(f_{\theta }({\mathbf{\Omega }}_{\mathbf{y}}))+\lambda {\hat{f}}_{\theta }(\mathbf{y})-(\lambda +1)\mathbf{y}{\Vert }_2^2\\ & +\eta \cdot \Vert h(f_{\theta }({\mathbf{\Omega }}_{\mathbf{y}}))-\mathbf{y}{\Vert }_2^2.\end{array}\end{array}$$

In the equation, $f_{\theta }(\cdot )$ represents any denoising network with an arbitrary architecture, while $h(\cdot )$ denotes a mask mapper capable of capturing global context. The term η is a fixed hyperparameter that controls the initial weight of the lossy blind denoising term and the stability of the training process. The variable λ is another hyperparameter that determines the strength of the visible (nonblind) denoising term when transitioning from the lossy blind phase to the lossless nonblind phase. The values of λ at the start and end of training are denoted by ${\lambda }_s$ and ${\lambda }_f$, respectively. To prevent the network from learning an identity mapping and to ensure that the denoising function $f_{\theta }(\cdot )$ actually removes noise, the gradient updates for $f_{\theta }(\cdot )$ are disabled when computing the gradients of the loss with respect to the clean target image $\mathbf{y}$. This is achieved by treating the output of $f_{\theta }(\mathbf{y})$ as a constant (denoted by ${\hat{f}}_{\theta }(\mathbf{y})$) that does not contribute to the gradient computation.

Image segmentation by the Swin-ResU network

The model employs a U-shaped network architecture, where the encoder’s feature representations are utilized in the decoder through skip connections at each resolution level. Patch merging layers are used at the end of each stage to downsample the resolution of the feature representations. In every resolution $i(i\in \{0,1,2,3,4\})$ of the encoder, the output feature representations are reshaped into $\frac{H}{2^i}\times \frac{W}{2^i}\times \frac{D}{2^i}$ and sent to the decoder blocks.

The encoder and decoder blocks are built upon the Swin Transformer architecture, which introduces 2 key modules: W-MSA and SW-MSA (Supplementary Fig. S2). W-MSA performs self-attention within local windows, significantly reducing the computational complexity compared to the standard MSA used in the Vision Transformer (ViT) (Dufter et al. 2022). Each Swin Transformer block consists of 2 subblocks: the first subblock uses W-MSA, while the second subblock employs SW-MSA. Each subblock is composed of a layernorm (LN) layer, an attention module (W-MSA or SW-MSA), another LN layer, and a multilayer perceptron (MLP) layer. This design allows for efficient local and global modeling of the feature representations (Liu et al. 2021).

The convolutional blocks in the decoder follow a similar structure to the encoder, with each block containing 2 consecutive 3 × 3 × 3 convolutional layers, followed by instance normalization and a Leaky ReLU activation function. The feature maps' resolution is then increased by a factor of 2 using a deconvolutional layer, and the outputs are concatenated with the corresponding outputs from the encoder via skip connections. These concatenated features are input into another residual block, as previously described. Finally, the segmentation outputs are computed using a 1 × 1 × 1 convolutional layer and a sigmoid activation function.

By incorporating the Swin Transformer blocks with W-MSA and SW-MSA into the U-shaped architecture, the model achieves a balance between local and global modeling while maintaining computational efficiency. The hierarchical feature representations learned by the encoder are effectively fused with the upsampled features in the decoder, enabling precise segmentation of the input data.

At present, the loss function utilizes a combination of cross-entropy and soft Dice loss, as illustrated in the equation:

$$\mathcal{L}(G,Y)=1-2{\displaystyle \frac{{\sum }_{n=1}^N{G}_n{Y}_n}{{\sum }_{n=1}^N{G}_n^2+{\sum }_{n=1}^N{Y}_n^2}-{\displaystyle \frac{1}{N}{\displaystyle \sum _{n=1}^N}{G}_n\log Y_n}}$$

In this equation, N denotes the number of voxels, while $G_n$ and $Y_n$ represent the probability output and corresponding ground truth at the Nth voxel.

We implemented our deep-learning framework using PyTorch Connectomics (Lin et al. 2021) and MONAI (Cardoso et al. 2022), and trained the model on 8 NVIDIA Tesla V100 32G GPUs. The training data consist of several consecutive sequence data randomly extracted from the entire data set, accounting for approximately 1% of the total, and annotated by experts. During training, random patches of 4 × 256 × 256 were cropped from 3D image volumes. For data augmentation, we employed techniques including the following:

1. Geometric transformations: image rotation and mirroring with 8 transformation directions, mirror padding and random cropping, and random elastic deformation. These augmentations help the model learn invariance to spatial transformations.

2. Color space augmentations: CutBlur, CutNoise, and Mixup. These techniques alter the color distribution and inject noise to increase model robustness to image quality variations.

3. Anomaly simulation: registration offset, motion blur simulation, partial loss of same-layer images (2D), and partial loss of sequence images (3D). By simulating realistic image anomalies, the model learns to handle imperfect data often encountered in real-world applications.

The model was trained for a total of more than 1,000 epochs with a linear warmup and using a cosine annealing learning rate scheduler. The initial learning rate is set to 0.001. For the loss function, we employed the combo loss, which combines weighted cross-entropy and dice loss, achieving better segmentation results on imbalanced data. To further improve the segmentation accuracy of small objects, we adjusted the foreground weights in the loss function. During inference, we used the sliding window approach, with an overlap rate of 0.5 for adjacent patches. This allows the model to make predictions on large volumes by stitching together the output probability maps.

Graph-based superpixel aggregation for 3D reconstruction

The cell wall probability map is oversegmented using a 2D distance transform watershed algorithm based on distance transformation (Meyer 1994). The specific steps are as follows: first, filter out small connected cell wall boundary components through thresholding operations. Then, considering the strong anisotropy of the data (horizontal resolution of 0.57 μm/pixel, axial resolution of 2 μm/pixel), we perform a Euclidean distance transformation on each cell wall probability map. After Gaussian smoothing, seed points are introduced at local distance maxima to allow executing a standard watershed process, thus obtaining 2D watershed superpixels. The superpixels generated by the distance transformation watershed are fragmented. In fact, the distribution of a complete cell often spans numerous superpixels, with each superpixel representing only a part of a cell. Therefore, the following steps involve merging superpixels to achieve 3D reconstruction of cells (Beier et al. 2017).

Images can be represented at a more abstract level through watershed superpixels, which involves constructing a 3D superpixel adjacency graph $G(V,E)$ with superpixels as vertices and the physical connections between superpixels as edges. The merging task for superpixels can then be represented as a graph partitioning problem on $G(V,E)$. Specifically, we model this problem as the MCMP (Beier et al. 2017):

$${\displaystyle \underset{y\in {\{0,1\}}^{|E|}}{min}}{\displaystyle \sum _{e\in E}}{c}_e{y}_e$$

$$s.t.y_e\le {\displaystyle \sum _{e^{\mathrm{\prime }}\in C}}{y}_{e^{\mathrm{\prime }}}\mathrm{\forall }e\in C\mathrm{\forall }C\in \mathrm{cycles}(G)$$

Here, $c_e\in \mathrm{\Re }$ represents the cost of separating the 2 endpoints of edge e into 2 distinct clusters, and $cycles(G)$ denotes the set of all cycles of graph G. The Boolean decision variables $y=\{y_e{\}}_{e\in E}$ indicate for each edge being cut (= 1) or uncut (= 0). In the original oversegmented image, cutting an edge implies preserving the boundary between the 2 corresponding superpixels, while merging an edge means that 2 superpixels belong to the same object.

We extract 3 types of hand-designed features: edge features of the boundaries between superpixels, edge features of the superpixels themselves, and interlayer features of anisotropic data. In this way, we design a 625-dimensional feature vector for each edge. Then, we train RF classifier to predict the probability of each edge e being cut,$p_e\in (0,1)$. Finally, using negative log-likelihood, we convert the probability values into the edge costs in the MCMP formulation, by:

$$c_e=\log \left({\displaystyle \frac{1-p_e}{p_e}}\right)$$

Thus, an MCMP instance is obtained.

For the given NP-hard multicut problem, exact solvers do not have polynomial time-solving efficiency. Hence, we treat this combinatorial optimization problem as a sequential decision process of edge contraction on a graph and use an approximate heuristic algorithm, GAEC (Keuper et al. 2015) (Supplementary Table S1), to obtain a suboptimal solution. The edge-based greedy algorithm iteratively updates solutions through edge contraction. The specific operation is to merge the 2 endpoints of the selected edge into a larger vertex and, at the same time, merge the edges connected to the 2 vertices into a larger edge with their weights summed accordingly (Beier et al. 2015). Finally, the obtained solution is mapped back to the original oversegmented image to get the final 3D reconstruction of the biological tissue, where each cell is assigned a unique label.

Benchmark testing experiments

The public data sets used in this study for evaluating the performance of Blind2Unblind are as follows: (i) Serial section Transmission Electron Microscopy (ssTEM) simulated data set—the CREMI data set (Funke et al. 2016); (ii) ATUM-SEM simulated data set—the widely recognized neuronal segmentation challenge data set SNEMI3D (Kasthuri et al. 2015); and (iii) ATUM-SEM real electron microscopy data—the Arabidopsis stamen data obtained by ATUM-SEM at the Institute of Automation, Chinese Academy of Sciences. The denoising network employs an enhanced U-Net architecture similar to that of Laine19 (Laine et al. 2019) and NBR2NBR (Ronneberger et al. 2015; Huang et al. 2021). The optimizer uses Adam with a weight decay coefficient set to 1e⁻⁸ to prevent overfitting. For both simulated and real electron microscopy denoising tasks, an initial learning rate of 0.001 was established. The learning rate is halved every 20 epochs within a total of 100 training epochs. Hyperparameters for the revisible loss were configured as η = 1, ${\lambda }_s$ s = 2, and ${\lambda }_f$ = 20, based on experimental experience. During training, EM data were randomly cropped into image patches of 128 × 128 pixels. In line with the N2V approach (Krull et al. 2019), the value of blind-spot pixels was set to be the weighted average of their 3 × 3 neighborhood pixels. All models were trained on a server using Python 3.8.5, PyTorch 1.7.1, and Nvidia Tesla V100 GPUs.

Quantitative evaluation metrics for denoising performance are PSNR and SSIM. PSNR calculates the ratio of the peak image signal to the mean squared error of image distortion. A higher PSNR value indicates better image quality restored by the denoising algorithm. It can be defined as:

$$\mathrm{PSNR}=10\cdot {\log }_{10}\left({\displaystyle \frac{{255}^2}{\mathrm{MSE}}}\right)$$

The SSIM is a metric for assessing the structural similarity between clean true-value electron microscopy images and denoised images. The closer the SSIM value is to 1, the more similar 2 images are, making it widely used for assessing image quality in image restoration and denoising tasks. It can be defined as:

$$\mathrm{SSIM}(I,I^{\mathrm{\prime }})={\displaystyle \frac{(2{\mu }_{\mathrm{I}}{\mu }_{{\mathrm{I}}^{\mathrm{\prime }}}+c_1)(2{\sigma }_{\mathrm{I}{\mathrm{I}}^{\mathrm{\prime }}}+c_2)}{({\mu }_{\mathrm{I}}^2+{\mu }_{{\mathrm{I}}^{\mathrm{\prime }}}^2+c_1)({\sigma }_{\mathrm{I}}^2+{\sigma }_{{\mathrm{I}}^{\mathrm{\prime }}}^2+c_2)}}$$

The public testing data set for evaluating the robustness and effectiveness of the Swin-ResU network, MSD Hepatic Vessel, can be found at Antonelli et al. (2022). And the BraTS21 (MR images) data set can be found at Baid et al. (2021). The comparison methods, including nnUNet, nnFormer, Genesis, and TransVW, can be found at Isensee et al. (2021), Liu et al. (2022), Shelhamer et al. (2017), and Haghighi et al. (2021). To facilitate the comparative experiments, we replaced the CNN decoder part of the Swin-ResU network with a transformer module, referred to as Swin-ResU-Trans, while the original Swin-ResU decoder was implemented using CNN, referred to as Swin-ResU-Cnn. Mean Dice score and HD are utilized as segmentation benchmarks.

From the LSFM data set with submicron resolution, patches of size 1192 × 3728 × 5 were taken every 50 layers as the test set, which includes a total of 5 groups. Ground-truth data have been annotated by experts for comparative analysis. The related methods are referenced in Hatamizadeh et al. (2022), Antonelli et al. (2022), and Baid et al. (2021). The algorithms for plant cell reconstruction, such as TWANG, Supervoxel-CNN, and Cellpose, have been implemented in the open-source software XPIWIT (Bartschat et al. 2016) (https://github.com/stegmaierj/XPIWITPipelines/tree/main). PlantSeg, which was proposed by Wolny et al. (2020), has been documented at https://bioimage.io/#/.

Assessment criteria for the efficiency of segmentation and aggregation

Precision is a measure of the proportion of true positive predictions among all positive predictions made by a model. It is calculated as:

$$\mathrm{Precision}={\displaystyle \frac{\mathrm{True}\mathrm{\setminus }\mathrm{Positives}}{\mathrm{True}\mathrm{\setminus }\mathrm{Positives}+\mathrm{False}\mathrm{\setminus }\mathrm{Positives}}}$$

Recall is a measure of the proportion of true positive predictions among all actual positive instances. It is calculated as:

$$\mathrm{Recall}={\displaystyle \frac{\mathrm{True}\mathrm{\setminus }\mathrm{Positives}}{\mathrm{True}\mathrm{\setminus }\mathrm{Positives}+\mathrm{False}\mathrm{\setminus }\mathrm{Negatives}}}$$

Accuracy is a measure of the proportion of correct predictions (both true positives and true negatives) among all predictions made by a model. It is calculated as:

$$\mathrm{Accuracy}={\displaystyle \frac{\mathrm{True}\mathrm{\setminus }\mathrm{Positives}+\mathrm{True}\mathrm{\setminus }\mathrm{Negatives}}{\begin{array}{c}\mathrm{True}\mathrm{\setminus }\mathrm{Positives}+\mathrm{False}\mathrm{\setminus }\mathrm{Positives}+\mathrm{True}\mathrm{\setminus }\mathrm{Negatives}\\ +\mathrm{False}\mathrm{\setminus }\mathrm{Negatives}\end{array}}}$$

Vinfo is calculated as$\mathrm{Vinfo}=V{I}_s+V{I}_m=H(A\mid Q)+H(Q\mid A)$; it indicates the amount of split errors. A means proposal and Q means ground truth. $\mathrm{ARAND}$ is calculated as $\mathrm{ARAND}=1-$ $\frac{2{\sum }_{ij}{o}_{ij}^2}{{\sum }_i{\left({\sum }_j{o}_{ij}\right)}^2+{\sum }_j{\left({\sum }_i{o}_{ij}\right)}^2}$; it indicates the amount of merge errors.

Clustering and cell type annotation

Through automated segmentation, we conducted an extensive analysis of segmented cells, extracting 80 morphological descriptors that referred to the Amira software (www.thermofisher.cn), including volume3d, area3d, elongation, sphericity, ellipticity oblate, and so on. The detailed description of these features was as follows: volume3d, volume of cells; area 3d, surface area of the cell boundary. Elongation is the ratio of the medium to the largest eigenvalue of the covariance matrix; elongated cells will have small values close to 0. Sphericity is the degree to which a cell approximates a spherical shape in its geometric form, and spherical cells have large values. Ellipticity oblate: two axis lengths of the ellipsoid were equal, and the different one was the shortest axis length; more oblate cells have large values close to 1. Ellipticity prolate: two axis lengths of the ellipsoid were equal, and the different one was the longest axis length; more prolate cells have large values close to 1. Additionally, detailed descriptions of all features can be found in Supplementary Data Set 5.

By referencing parameter ranges and comparing with segmented results from X-ray images, we filtered out segmentation errors. Data were downsampled to a resolution of 80 × 80 × 100 nm³ to compute all features. This resulted in a table of 80 features encompassing segmented cells for P. canadensis Moench. and P. simonii Carr., respectively. Using Seurat version 4.3.0 (Hao et al. 2021) (https://github.com/satijalab/seurat), the data were standardized by the z-score method, and principal component analysis (PCA) and UMAP were performed. We utilized the K-nearest neighbor (KNN) algorithm to cluster the cells with a resolution parameter of 0.3 and 0.5 for P. canadensis Moench. and P. simonii Carr. seeds, respectively. Finally, we generated the UMAP embedding using the RunUMAP function with principal components and parameters n_neighbors = 10 and min_dist = 0.1. To investigate the morphological differences among different cell types, we defined cells into 7 types based on spatial location information, including Ep cells, Co cells, columella cells, radical cells, SAM cells, and Hy cells.

To cluster the SAM cells from seedlings, we employed the identical methodology as described above, applying a resolution parameter set at 0.3. After that, the 3D mapping was constructed using the R package plotly, version 4.10.4.

Differentiation analysis of cell shape

To compare the cytological features of cell types between poplar seeds, we conducted the Mann–Whitney–Wilcoxon test for all cytological features using the wilcox.test function (parament: paired = FALSE) in R-base software package (version 4.2.0). We then calculated the P values representing the differences respectively and generated heatmaps using the pheatmap package (version 1.0.12). To compare and visualize the shape differences in cells between 2 independent Cos, we used Imaris 9.6 to perform 2D rendering images of the cell area and ellipticity prolate, and the local density of Co cells was calculated and rendered using SR-Tesseler software (http://www.iins.u-bordeaux.fr/teamsibarita-SR-Tesseler).

Topological analysis

The gaps, contours, and cell walls are identified by the network in the second phase of this framework. In the third step of the cell aggregation process, the boundary information will be utilized. The acquisition of cell connectivity is performed after the aggregation process, when the complete reconstruction result of the tissue is obtained. At this point, each individual cell has been fully segmented and assigned a unique global label. After that, the contact area and contact network are calculated by the python package, connected-components-3d (https://github.com/seung-lab/connected-components-3d). It can accurately compute the contact surface area and generate a voxel connectivity graph. During the process, only face contacts are considered edges and corners. Subsequently, the boundary information obtained from the cell wall results in the second step is applied to the aggregated results, resulting in the final reconstructed volume with cell walls, contours (gaps), and other boundary information. After obtaining the contact surface area of all cells in each tissue, use the open-source software Cytoscape_v3.9.1 (http://www.cytoscape.org/) to analyze the degree and BC of cells. Specifically, the term “degree” refers to the number of neighboring nodes a node has. The term “shortest path” refers to the minimum number of distinct edges that pass through from one vertex to another in the network (Barthélemy 2011). And the term “BC” refers to the frequency at which a node lies on the shortest paths between all other nodes. Use the minimum–maximum normalization method to normalize the BC data for comparison. Finally, the results were analyzed and visualized by GraphPad Prism version 9.0.0.

Semithin sections of the P. simonii Carr. seed and light microscopy observation

The mature seeds were fixed in fixative (2.4% paraformaldehyde and 2% glutaraldehyde) for 2 h under vacuum and then incubated at 4 °C overnight. After washing 4 times with 0.1 m PBS buffer, the seeds were then sequentially transferred to increasing percentages of ethanol solution (30%, 50%, 70%, 80%, 90%, 95%, 3 × 100%) and different ratios of 100% ethanol and acetone (1:2, 1:1, 2:1, 100% acetone) series to dehydrate for 30 min, respectively. Then, the seeds were infiltrated with a series of increasing concentrations of Spurr resin (Electron Microscopy Sciences, in the ratio of 10 g ERL 4221, 6 g DER 736, 26 g NSA, 0.4 g DMAE). Finally, the samples were embedded in polymerized resin using an embedding mold for 12 h at 45 °C and 48 h at 70 °C. The embedded samples were sliced horizontally or vertically with a fully automated rotary microtome Leica RM2265, stained with toluidine blue, and finally observed by Leica S9i microscope.

Author contributions

The experiments were designed by Jin.L., X.Z., and H.H. Z.H. and S.S. contributed to collecting the X-ray and LSFM images. Jia.L., Z.H., and J.Y. contributed to developing the high-precision automated method for segmenting all type cells. S.S. and Z.H. contributed to analyzing 3D data of segmented cells. Z.H., Jia.L., and S.S. wrote the manuscript. W.W., W.S., G.W., L.M., S.Y., Y.C., Z.L., L.S., L.L., J.B., and X.X. helped in conducting the 3D reconstruction. All authors participated in the review of the manuscript.

Supplementary data

The following materials are available in the online version of this article.

Supplementary Figure S1. Graphical flowchart of X-ray imaging and analysis.

Supplementary Figure S2. Details of the global perception mask mapper and the 3D Shifted Windows Multi-head Self-Attention (SW-MSA).

Supplementary Figure S3. Comparative and visualized results of different denoising methods processed based on the SNEMI3D test set (Gaussian noise σ = 25).

Supplementary Figure S4. Comparative and visualized results of different denoising methods processed based on the SNEMI3D test set (Poisson noise λ = 10).

Supplementary Figure S5. Comparative and visualized results of different denoising methods processed based on the real electron microscopy data of the Arabidopsis stamen by ATUM-SEM.

Supplementary Figure S6. Comparison of segmentation and aggregation performance of our approach (Swin-ResU + Multicut) on the LSFM images under different denoising methods.

Supplementary Figure S7. Comparison of the effect for cell segmentation and aggregation based on different algorithms.

Supplementary Figure S8. Benchmark testing and comparative results of our approach and existing methods.

Supplementary Figure S9. The feature plots and violin distributions of 4 features in segmented cells from different seeds.

Supplementary Figure S10. Proportion of different cell types relative to each cluster in P. simonii Carr. and P. canadensis Moench.

Supplementary Figure S11. Comparison of features and significant differences (P values) between P. canadensis Moench. and P. simonii Carr. cells.

Supplementary Figure S12. Comparative analysis of geometric morphology and cell features of bilateral Cos in P. canadensis Moench. and P. simonii Carr. seeds.

Supplementary Figure S13. The effect of LVACR on cell identification in SAM tissues and young leaves of pine seedlings.

Supplementary Figure S14. Dynamic changes in the numbers of SAM cells at different developmental stages.

Supplementary Figure S15. Dynamic changes and visualization of the 2D cell area of SAM cells at different developmental stages.

Supplementary Table S1. Edge-based greedy algorithm process.

Supplementary Table S2. Quantitative evaluation of static and dynamic Gaussian noise on simulation data sets.

Supplementary Table S3. Quantitative evaluation of static and dynamic Poisson noise on simulation data sets.

Supplementary Table S4. Benchmark testing using MSD data set.

Supplementary Table S5. Benchmark testing using BraTS21 data set.

Supplementary Data Set 1. Detailed feature values of segmented cells in different tissues.

Supplementary Data Set 2. Quantitative evaluation of the efficiency of denoising for 5 different data sets based on the Blind2Unblind, Gaussian, or no denoising.

Supplementary Data Set 3. Quantitative evaluation of the efficiency of segmentation for 5 different data sets based on different segmentation networks and aggregation algorithms.

Supplementary Data Set 4. LSFM data set segmentation benchmarks in terms of precision, recall, accuracy, Vinfo, and so on.

Supplementary Data Set 5. Description of 80 features used in clustering analysis.

Supplementary Data Set 6. All detailed values of features of segmented and filtered cells in P. canadensis Moench. and P. simonii Carr. seeds.

Supplementary Data Set 7. Detailed topological values of segmented cells in different poplar seeds.

Supplementary Data Set 8. Detailed values of morphological features of bilateral Cos in P. canadensis Moench. and P. simonii Carr. seeds.

Supplementary Data Set 9. All detailed values of features of segmented and filtered cells at different stages.

Funding

This work was supported by grants from the National Natural Science Foundation of China (32030010), STI 2030-Major Projects (2022ZD0401605), Fundamental Research Funds for the Central Universities (QNTD202301), Beijing Municipal Natural Science Foundation (5232016), the National Natural Science Foundation of China (32000558, 32171461, and 32370740), Beijing Nova Program (20230484251), the Program of Introducing Talents of Discipline to Universities (111 project, B13007), Scientific Research Instrument and Equipment Development Project of Chinese Academy of Sciences (YJKYYQ20210022), 5.5 Engineering Research & Innovation Team Project of Beijing Forestry University (BLRC2023C06), and Biological Breeding-National Science and Technology Major Project (2023ZD04069).

Data availability

The data underlying this article are available in the article and in its online supplementary material.