Connectomics, emerging from large animal datasets and limited human neural tissues, has expanded our understanding of brain disorders by revealing neural architectures.1^,2 Sellular-level micro-connectomics predominantly uses non-human models to reveal phenomena that extend to larger-scale behaviors in humans.3 Although advances in electron microscopy and mesoscale data acquisition have facilitated axon-level insights,4^,5 integrating confocal Z-stack imaging and 3D layer interpolation offers new potential to uncover cellular intricacies.6

Confocal microscopy has been widely used for morphological characterization to enable detailed exploration of biological structures, including surface and internal tissue configurations.7^–12 This technique is indispensable in organoid research to aid the study of neural development, nerve fibers, and brain structures.13^–24 Despite these benefits, confocal microscopy faces challenges, such as diffraction limits,25 sample damages due to photobleaching, phototoxicity,26 and time-consuming image acquisition processes.7

Although confocal microscopy remains a powerful tool, rapid and detailed 3D modeling of organoids presents difficulties, especially when dealing with complex neuronal structures prone to phenotypic variations.27 Artificial intelligence can be used to reduce reconstruction complexity and increase flexibility.28^–32 Deep-Z33 and other AI-driven schemes, such as GAN-based models,34 have been shown to improve 3D imaging by eliminating mechanical scanning and enhancing image resolution.35^–37 However, these methods often suffer from challenges in data scarcity and the computational complexity of reconstructing 3D volumes during real-time acquisitions.38^,39

Here, we introduce PLayer, a compact neural system designed for fast 3D volume reconstruction of organoids. By utilizing sparse input layers, PLayer avoids traditional 3D convolution techniques. Instead, it employs 2D convolution-based statistical map estimation (SME) to enhance features in neighboring layers. This is followed by inverse proportional function (IPF) interpolation, cascaded with a 2D Vision Transformer, to further refine and blend the features for super-resolved 3D volumes. With reconstruction times of around 20 seconds on a GPU, PLayer significantly accelerates the confocal microscopy workflow to offer high efficiencies in micro-connectomics research by addressing the key problems of missing layer information and extended acquisition time.

A Nikon A1+ modular laser scanning confocal microscope system, equipped with photomultiplier tubes (PMTs), was used for confocal imaging to acquire Z-stack images. The beam path is shown on the right side of Fig. 1. Objectives with 10×, 20× or 60× magnification were used, and images were captured at a resolution of $2048\text{pixel}\times 2048\text{pixel}$. The 405-, 488-, 561-, and 647-nm laser lines excited the selected fluorophores corresponding to DAPI, FITC, TRITC, and Cy5, respectively. The diffraction formula is as follows:$$R_{\text{axial}}=\frac{1.4\lambda \eta }{(NA{)}^2},$$(1)where $R_{\text{axial}}$ represents the axial resolution, $\lambda$ denotes the wavelength of the emission light, $\eta$ is the mounting medium’s refractive index, and NA is the numerical aperture of the objective.40 The calculated axial resolution of confocal microscopy was $\sim 500\mathrm{nm}$, which was set as the vertical step size in our microscopy setup.41

The high-voltage setting for the gallium arsenide phosphide photomultiplier tube (PMT) detector was adjusted between 31 and 80 during the acquisition process. The offset setting, which controlled the image’s background level by adjusting the PMT offset voltage, was set to 0. The power of the 488-nm laser used for excitation ranged from 11 to 24.60 mW. The powers for lasers corresponding to 405, 561, and 647 nm were 55, 20, and 20, respectively. The thickness of the slide imaging was set between 8.00 and $29.00\mu \mathrm{m}$, with each imaging step being 1, 0.5, and $0.1\mu \mathrm{m}$. This resulted in a specific quantity of raw images (e.g., 21 images for a $10\text{-}\mu \mathrm{m}$ thickness at $0.5\text{-}\mu \mathrm{m}$ steps). Although axial resolution could not exceed 500 nm due to wavelength limitations and pinhole design, we still imaged some slides at steps of 0.5 and $0.1\mu \mathrm{m}$ as permitted by the confocal imaging software system.

For each layer, the imaging process took roughly 7.7 s. With 21 layers per organoid, the time required for image acquisition ranged from a minimum of $\sim 2.47\min$ when one channel was selected to a maximum of $\sim 28\min$ for three channels at 60× magnification. For a sample with DAPI, TRITC, and Cy5 with 20× magnification and $0.5\text{-}\mu \mathrm{m}$ increments, it took $\sim 22.72\mathrm{s}$ to image one layer. Therefore, the total imaging time was 477.18 s or $\sim 7.95\min$.

Our method could be divided into four phases. As depicted in phase I of Fig. 2(a), the initial processing of images involved denoising through a GAN-based neural network. This step was crucial for enhancing learning efficiency.40^,42 Although prior knowledge indicated that Gaussian noise was the primary noise in confocal microscopy, the unknown distribution of the noise hindered filtering-based denoising. To mitigate spatial noise on the organoid layer without knowledge of the ground truth, we trained the GAN-based neural network to restore the organoid layer with additional Gaussian noise applied to the corresponding input. Subsequently, we employed the trained GAN-based neural network to process the raw images of the organoid layers, reducing this type of noise and enhancing the quality of the images of organoid layers. The GAN-based denoising model was trained with varying levels of Gaussian and Poisson noise as part of data augmentation.43 This approach helped the model generalize to the different noise types commonly found in fluorescence imaging.44

Figure 2(a) illustrates the PLayer algorithm. In phase II, we selected the four nearest neighboring layers as an example (from ${\text{Layer}}_{n-2t}$ to ${\text{Layer}}_{n+2t}$, where $t$ denotes the layer picking interval) to reconstruct and present the high-resolution ${\text{Layer}}_n$. It is also possible to select six or eight nearest neighboring layers to reconstruct the target layer. In the cases where neighboring layers were absent, for instance, the top 2 or bottom 2 layers of the organoid, a blank layer served as a substitute for the missing layer.

Before the Restorer stage, we first processed the neighboring layers with SME and then performed the interpolation with IPF to make the first guess of the target layer, which served as the feature extraction module to improve the efficiency of the IPF-based interpolation, as shown in Fig. 2(b). The details of SMEs are elaborated in Appendix A.4 in the Supplementary Material. In this phase, interpolation was employed, utilizing SME on the sample data to encode the interpolated ${\text{Layer}}_n$ in low-resolution and transmit it to phase III (Restorer) of PLayer for reconstructing the image of ${\text{Layer}}_n$. The Restorer consists of a feature-rich component and a UNet-like SR section, as illustrated in Fig. 2(a). Transformer blocks are utilized at each level of the UNet-like SR section to extract features corresponding to each level. By utilizing SME to extract features from neighboring layers, followed by interpolation, we eliminate the need for conventional 3D convolution operations in the Restorer.45 The details of the neural network are illustrated in Appendix A.3 in the Supplementary Material.

We used Restorer to directly process the 3D volume ($H\times W\times L=2048\times 2048\times L$, where $L$ was the layer number of the volume, and $H$ and $W$ have their units in pixels) to determine the high-resolution version of the target layer. As most fluorescence images are uneven in brightness in deep layers, we use the YUV color space instead of RGB to represent images in phase III. This allows the model to emphasize local contrast enhancement and dynamic range compression, improving its ability to perceive and process images during training. In addition, we introduce image augmentation methods that vary the $Y$ (brightness) channel to restore the original brightness, resulting in better training outcomes. As shown in process III in Fig. 2, we separately extract features from the $Y$ (brightness) channel and the UV (color) channels to specify the restoration of both color and brightness in process III-2, where a ViT-based UNet model is employed.

Compared with direct processing, it was more effective to employ a trainable neural network to encode low-resolution interpolated images, followed by decoding with Restorer. This approach not only reduced redundant noise during reconstruction but also optimized memory usage by reducing the data size of the 3D volume. In addition, we found that the encoding functions resembled a Masked Autoencoder46 operating at the pixel level. This feature proved beneficial in enhancing the reconstruction during the decoding process.

In post-training, the Restorer model enhanced the axial resolution of the organoid volume. The resulting images were incorporated into the volume and stacked in the z-dimension, forming a refined 3D organoid volume, visualized using the Visualization Toolkit and Mayavi package.47^,48 Finally, phase IV involved optimizing and deploying the model to balance accuracy and efficiency, ensuring its suitability for practical applications. Figure 2(c) illustrates the iterative pruning process, where the pruning ratio is adjusted based on accuracy evaluations. This approach removes small weights, enabling efficient deployment and reliable performance on embedded devices.49 Notably, the initial model may occasionally overfit the training dataset, and pruning can improve reconstruction accuracy on the test dataset. However, excessively increasing the pruning ratio can degrade performance. To address this, we established an expected accuracy threshold (Exp.), allowing the model to be pruned as much as possible while maintaining the desired reconstruction accuracy.

Training a Restorer for larger resolution improvement is formidably heavy in computation power and memory. To enable resolution enhancement for any desired factor, we reconstructed the middle layer between the top and bottom layers.50 As exemplified in Fig. 3(a), the algorithm started by setting a hyperparameter $n$, representing the desired factor for the compression-magnification resolution improvement (e.g., $n=2,4,8$, etc.). The SME method was employed to ascertain pixel probabilities on the target layer based on information from neighboring layers. Subsequently, the target layer was derived through interpolation of these neighboring layers. This interpolated image underwent a Restorer process for refining image quality. Finally, each iteration within the loop decremented $n$ by 1, ultimately revealing the middle layer of the 3D volume.

After the iteration ($n=n-1$), images were stacked together, culminating in a new and improved 3D volume. This iterative loop persisted until $n$ equals zero, to form the final super-resolution 3D volume. Figure 3(b) provides a visual representation of the interpolation process, in which different colors correspond to distinct interpolation phases. The reconstructed 3D volumes for each phase, presented on the right side of Fig. 3(b), offer a visual insight into the algorithm’s efficacy at each stage.

The organoid, with axons stained using the type IV neurofilament heavy antibody,51 served as the subject of this investigation. We imaged new 3D organoid volumes with confocal microscopy to gather a training dataset and deployed PLayer to reconstruct the selected 2D target layers in the volumes, allowing for performance evaluation at both the interpolation and reconstruction stages.

To demonstrate the impact of these interpolation methods within a neural network framework, a 3D neural organoid was randomly selected, and three segments were displayed in Figs. 4(a)–4(c). The process involved using odd layers (e.g., 1, 3, 5, …) as inputs to double the vertical resolution and extrapolate the full 3D organoid structure. Given the microscopy system’s spatial resolution limit of $0.5\mu \mathrm{m}$ between adjacent layers, gaps such as $1\mu \mathrm{m}$ (e.g., between ${\text{Layer}}_{n-3}$ and ${\text{Layer}}_{n-1}$) were filled through interpolation. These interpolated layers served as inputs to the Restorer, and the outputs were evaluated against reference images layer by layer. Figures 4(a)–4(c) organize the visualization into seven columns, each representing a distinct step of the process. Columns 1 to 4 show images from neighboring layers for two layers below (${\text{Layer}}_{n-3}$ and ${\text{Layer}}_{n-1}$) to two layers above (${\text{Layer}}_{n+1}$ and ${\text{Layer}}_{n+3}$) the target layer. Column 5 shows the ground truth of the target ${\text{Layer}}_n$, and column 6 shows the interpolated estimation of the target layer based on neighboring layers using interpolation—Interpolated ${\text{Layer}}_n$. Column 7 displays the Reconstructed ${\text{Layer}}_n$ after refinement by the Restorer. The 3D volume resolution is $2048\times 2048\times L$ ($H\times W\times L$, where $L$ is the layer number of the volume, and $H$ and $W$ have their units in pixels).

Our investigation explores methods for estimating the initial distribution of unknown layers through interpolation, including generalized normal distribution (GND),52 IPF, and traditional cubic and bilinear interpolation techniques, as shown in Figs. 4(d)–4(g). Both GND and IPF share a key feature of assigning larger weights to nearby layers and smaller weights to distant ones, but they differ in complexity: IPF uses a function with fewer parameters [equation (2) in the Supplementary Material], whereas GND employs a six-parameter function that poses challenges for generalization.52

The metrics used for evaluation, peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM), were computed using RGB images to reflect the multichannel markers in the organoid images, where color information plays a key role in the analysis. As a result, the calculated values are relatively low compared with those obtained from grayscale images. The hyperparameters used for training PLayer are detailed in Appendix A.1 in the Supplementary Material, where Table S1 lists the hyperparameters used in our experiments. The analysis in Figs. 4(d) and 4(e) highlights the superior performance, stability, and optimal results of IPF in both PSNR and SSIM metrics, with significant advantages in extrapolating intermediate layer structures to outperform other interpolation methods.

Figures 4(f) and 4(g) highlight the significance of using IPF as the interpolation method. We compared various interpolation techniques, including GND, cubic, bilinear, and IPF, and subsequently performed reconstruction using the Restorer. The results demonstrate that IPF-based interpolation yields superior outcomes compared with the other methods, indicating that the exponential-function-like interpolation provided by IPF is more effective in generating an initial estimate of the target layer than other interpolation approaches. All images were preprocessed using the GAN-based denoising method as detailed in Appendix A.2 in the Supplementary Material, where Fig. S1 presents a comparative analysis of various image denoising methods. The $x$-axis in subplots (f)–(g) represents the layer number along the axial direction of the organoid, showing how denoising improves consistency across layers.

One function of the Restorer in this phase was to reconstruct the 2D organoid fiber from the most reliable SME information while removing artifacts introduced by interpolation. This reconstruction step, performed using a neural network trained through transfer learning (first using MRI data,53 followed by constrained neural organoid confocal $Z$-stack images), removed artifacts and enhanced resolution. This method not only achieved clearer outputs but also reduced GPU memory requirements, enabling the deployment on embedded devices. For boundary layers lacking sufficient neighbors (e.g., ${\mathrm{Layer}}_{n-3}$ at the top or ${\mathrm{Layer}}_{n+3}$ at the bottom), blank images were used as placeholders to maintain consistency, as seen in Figs. 4(b) and 4(c).

Refinement by the Restorer significantly improved PSNR and SSIM values compared with interpolation alone, as depicted in Figs. 4(h) and 4(i). Beyond numerical metrics, the reconstructed layers exhibited enhanced structural fidelity, with organoid fiber composition more closely resembling the ground truth. This improvement is evident in the white arrows in Figs. 4(b) and 4(c), where the reconstructed fibers more accurately represent the original target layer structures.

We define reconstruction time as the total time required to complete the entire reconstruction process, including both image upload (data loading) and the computational reconstruction of missing layers. The processing time refers specifically to the time spent on the computational steps involved in reconstructing the intermediate layers. In our experiment, factors of 8× are faster than those of 4× and 2× for the overall reconstruction because it requires fewer input layers, thereby reducing data transfer overhead. Conversely, the increased number of input layers in the 4× and 2× methods leads to longer data-loading times due to GPU and CPU memory allocation constraints. The 3D organoid reconstruction of the PLayer result is depicted in Fig. 5. The raw 3D volume with a vertical resolution of $0.5\mu \mathrm{m}$, as shown in Fig. 5(a), was denoised and transformed into a low-resolution (LR) 3D volume with an axial resolution of $0.5\mu \mathrm{m}\times t$ in Fig. 5(b) by downsampling every $t$ layer. Here, we set $t=2$ as an example. We first interpolated the downsampled LR 3D volume in Fig. 5(b) using our interpolation method, into a 3D volume with an axial resolution of $0.5\mu \mathrm{m}$ in Fig. 5(c), whereas there were still discernible artifacts in Fig. 5(c).

The results of Restorer processing without (w/o) SME and with SME are illustrated in Figs. 5(d) and 5(e), respectively. The reference in Fig. 5(a) is the original 3D volume without post-processing. After being processed by the Restorer, the encoded layers were reconstructed, yielding an SR-3D volume with clear fibers, as depicted in Figs. 5(d) and 5(e). The artifact suppression of SME in the reconstruction process is highlighted by the white arrows in columns Fig. 5(d) versus Fig. 5(e). The first and second rows of the five columns represent the 3D volume from a solid angle with azimuth = 0, elevation = 90, roll = 0, and azimuth = 60, elevation = 90, roll = 0, respectively. Figure 5(f) shows the 3D volume of SR-3D without SME with the solid angle of azimuth = 30, elevation = 90, and roll = 0. By contrast, Fig. 5(g) presents the 3D volume of SR-3D with SME under the same angle.

The enlarged purple rectangle pair between Figs. 5(f) and 5(g) emphasizes the artifact-suppressing ability of SR-3D reconstruction without SME versus with SME, as highlighted by the yellow arrows. Figure 5(h) displays the quantitative outcomes of 3D reconstruction evaluated using the NIQE, proving the superiority of super-resolution 3D reconstruction using SME. The SR-SME method exhibits superior NIQE compared with those of standard SR images, and both of them are better than interpolated results. Here, the relatively poor quality index of the reference images can be attributed to the artifacts in the organoid fibers. The quantitative assessment reinforces the effectiveness of our approach in achieving high-quality 3D reconstruction with reduced artifacts. Video 1 shows the comparison of 3D volume reconstruction with PLayer against high resolution (HR) and LR volumes.

To demonstrate how well our method reconstructs 3D organoid volumes with undersampling ranging from the compression-magnification factors of 2× to 8× along the axial direction, we trained PLayer to handle axial resolutions at 2×, 4×, and 8× super-resolution ratios. Figure 6 shows the results: the first row is a randomly selected 2D layer from the 3D volume, whereas the second to fourth rows display the reconstructions at different levels of axial resolution. As the undersampling increases from 2× to 8×, some fine details become less visible (highlighted by yellow square), though regions with dense organoid fibers (highlighted by white square) remain well-preserved.

Figures 6(a) and 6(b) quantify PSNR and SSIM at different layers of the 3D reconstruction, with different undersampling ratios of the axial resolution. We can observe a slight deterioration of these metrics at a higher undersampling ratio. The periods of 2×, 4×, and 8× metric variation overlap on layers 8 and 16 (data index) because these layers were known inputs of the model. Degraded reconstruction quality is observed in the first layer of enhancements at 2×, 4×, and 8× compression factors. This degradation is attributed to insufficient GPU utilization during the initial stages of inference, where computational overheads such as initialization and data transfer dominate the process. These inefficiencies lead to suboptimal calculations and compromise the accuracy of the initial layer’s reconstruction. To demonstrate these variations in GPU performance, we model a warm-up phase, which is further discussed in Appendix A.5; Table S2 in the Supplementary Material). Figures 6(c) and 6(d) provide insights into the reconstruction time (end-to-end pipeline duration) required by different models with GPU and CPU, respectively. In Fig. 6(c), we observe that when reconstructions are performed on GPUs, the times for 2×, 4×, and 8× enhancement levels are nearly equivalent during inference. To show the time consumption clearly, we use the CPU to do the reconstruction. Figure 6(d) illustrates the reconstruction time required to achieve an 8× improvement in axial resolution. Although 8× super-resolution involves increased model complexity, we observed that it requires fewer input layers and less overall reconstruction time when compared with those of the 4× and 2× processes. Although the 8× super-resolution demands more processing time (per-image inference time), it benefits from reduced time spent loading the input layers of the organoid to ultimately reduce the computational load. This is because our approach utilizes neighboring layers to predict the target layer, which requires more data input as compared to those of traditional 2D images with super-resolution. As a result, our experiments showed that 8× super-resolution was faster than those of both 4× and 2× processes, primarily due to the reduced time spent on image loading. Therefore, the analysis indicates that the primary limitation lies in input/output image processing as more time is allocated to processing input layers rather than that of the intermediate layer computation. As such, fewer input layers lead to faster reconstruction time.

The full potential of PLayer is realized when deployed on embedded devices. In this study, we demonstrate the 2D and 3D interpolation and reconstruction of neural organoids using PLayer on both an Nvidia GPU and a Raspberry Pi. Restorer, based on the Vision Transformer, has a model size of approximately 19 MB, which does not significantly impact inference time on a GPU. We employed model pruning and quantization methods of embedded neural network (ENN) to alleviate the computational burden54 and resize the input images to save processing time, denoting it as a reside-embedded neural network (R-ENN).

Figure 7 compares model performance on GPU and Raspberry Pi, where displaying a $2048\times 2048\times 20$ volume on Raspberry Pi required at least 64 MB of memory. The reconstruction process, which relies on a pre-trained model, further consumes memory, causing the system to quickly run out of resources when handling high-resolution images. Consequently, we had to down-sample 2D layers from $2048\times 2048$ to $1024\times 1024$. Figures 7(a)–7(c) provide comparisons of performance metrics for reconstructing 5 3D volumes with a resolution of $1024\times 1024\times 20$ ($H\times W\times L$, where $L$ represents the number of layers, and $H$ and $W$ have their units in pixels). The reconstruction methods evaluated include CNN on GPU, CNN on Pi, ENN on Pi, and R-ENN on Pi. The 100 samples on the $x$-axis correspond to 100 randomly selected 2D layers from our test dataset, which consists of eight organoid volumes, each containing 21 layers. The evaluation metrics include SSIM, PSNR, and reconstruction time. In this example [Fig. 7(a)], the average SSIM values for CNN and ENN on both GPU and Pi ranged from 0.90 to 0.92. In addition, as depicted in Fig. 7(b), the average PSNR values for CNN and ENN on both GPU and Pi ranged from 28 to 31 dB. For R-ENN, the SSIM values and the PSNR for each layer were lower than those of other approaches. As we downsampled the R-ENN input and upsampled the output into a 3D volume, R-ENN’s SSIM and PSNR are lower than others with much faster reconstruction time.

In both embedded and nonembedded neural networks, the values for both SSIM and PSNR proved satisfactory. However, a more substantial discrepancy was in the layer reconstruction time. CNN on Pi and ENN with the pruning ratio of 21.83% computed with Pi exhibited the longest processing times for each layer, at $\sim 30\mathrm{s}$. Meanwhile, CNN on GPU and R-ENN on Pi performed comparably with a much shorter reconstruction time of $\sim 2\mathrm{s}$. This highlights the effectiveness of R-ENN in reducing computational complexity (Appendix A.6 in the Supplementary Material). Figure. S2 in the Supplementary Material presents a demonstration of displaying the 3D volume of the organoid on an embedded device.

It is observed that the desired pruning ratio often differs from the actual pruning ratio achieved by the model. For instance, in our experiments, when aiming to prune 80% of the parameters, the model only pruned 10.24% of the nonzero parameters. This discrepancy could be attributed to the presence of pre-existing zero parameters in the model. When we specified a pruning ratio of 80%, the model first accounted for the number of zero parameters already present and then attempted to prune additional nonzero parameters to meet the target. To strike a balance between efficiency and accuracy, we investigated various pruning ratios along a spectrum.49^,55Figures 7(d)–7(f) reveal the performance of the ENN model in reconstructing the target layer with a resolution of $1024\times 1024$ across different pruning ratios, ranging from 10.25% to 76.71%. The optimal SSIM, as depicted in Fig. 7(d), and PSNR, as observed in Fig. 7(e), appeared at a pruning ratio of 10.25%. Compared with the results of the pruning ratio of 0% (CNN computed on Pi) in Figs. 7(a) and 7(b), we observed an increase in both PSNR and SSIM appeared when the pruning ratio is 10.25%. This phenomenon arose from the fact that the original unpruned model was overfitted on the small training dataset of 3D organoid volumes. By pruning certain small weights of the model, we enhanced its suitability for the test dataset and fostered the model’s generalization. However, as the pruning ratio progressed towards 80%, the diminution of PSNR and SSIM indicated the model’s degradation caused by over-pruning.

Figures 7(g)–7(j) reveal the visualized 3D volume reconstruction achieved through distinct methods. The white arrows pinpoint the disparities between results derived from R-ENN and others. Although Figs. 7(g)–7(i) exhibits almost identical reconstruction levels, the fastest R-ENN model on Pi produced a comparatively less optimal visual result, underscoring the inherent trade-off between 3D volume reconstruction quality and speed in the R-ENN model. Further comparisons among the inference results using a traditional neural network on GPU, ENN, and R-ENN on the Raspberry Pi are presented in Video 2. Furthermore, we show the deployed embedded neural network on Raspberry Pi in Video 3.

Using neural organoid confocal data for preprocessing, training, and 3D reconstruction, PLayer is vital for understanding biological patterns. Organoids, characterized by abundant, chaotic nerve fibers and fascicles, display meaningful connections in 3D space. We selected organoids of different ages, generated through unguided protocols56 mainly for training and regionalized neural organoid protocols57 predominantly for testing. Our antibodies for staining included neuron-specific class III $\beta$-tubulin (TUJ1) and type IV neurofilament heavy (NF) to create two distinct scenarios.

TUJ1 is used to identify young neurons, including newly formed and differentiated cells.58^,59 Given the high expression of young neurons in neural organoids,56 we ensured confocal images produced vibrant visuals with dense nerve fiber concentrations. Background noise and class imbalance posed challenges for neural networks during training,60 affecting convergence,61classification efficacy,62 and network generalization.63 However, these data strengthened PLayer, allowing us to demonstrate denoising methods (Appendix A.7 in the Supplementary Material) for mitigating autofluorescence issues. Table S3 in the Supplementary Material presents a comparison between PLayer and other 3D reconstruction methods across different evaluation metrics.64

We also exposed our algorithm to scenarios with low brightness, using NF antibody, prevalent in mature myelinated axons.51 Our day 82 NF antibody-stained dorsal neural organoids57 effectively modeled continuous fiber spans.

PLayer aids micro-connectomics research by reconstructing consistent, continuous, and accurate image layers axially from minimal input to compensate for biological sample-associated imaging degradation. With only two images of a 3D sample, PLayer reconstructs additional layers not originally captured and shortens the imaging process time compared with those of traditional 3D reconstruction methods.65^–69 Model pruning and image resizing can further enhance PLayer’s efficiency.

Despite advancements in deep learning for 3D microscopy, confocal microscopy for organoid volume reconstruction to uncover micro-connectomics remains underexplored. Previous works included 3D refocusing of fluorescence microscopy with deep learning33 and snapshot imaging methods,36 which reconstructed only the 3D surface. By contrast, our method leverages learning from neighboring layers to stack layers into a volume and reconstruct 3D images. Our approach also mitigates artifact migration, contrasting with the unsupervised Cycle-GAN method35 that introduced visual artifacts. We incorporate SME to address this issue, along with preprocessing steps.

In addition, our method excels in computing resource efficiency. Prior medical image 3D reconstruction methods often relied on complex convolutions, making them computationally intensive. Examples include a 3D GAN for wide-field microscopy and a diffusion model for MRI denoising.34^,70 By contrast, PLayer uses a 2D embedded neural network design that accelerates reconstruction and prevents GPU memory depletion, a common challenge with 3D volume processing. Previous works in 3D single-cell imaging analysis faced limitations in annotation and datasets,37 hindering real-time inference. Our approach utilizes pruning techniques to enhance model efficiency and effectiveness.

Rather than using the original interpolated layer directly, PLayer employs a downsampled 3D volume for acceleration and improved memory efficiency (Appendix A.8 in the Supplementary Material) supported by findings in FifNet. Table S4 compares the performance of PLayer with other methods in terms of memory usage, time efficiency, and reconstruction accuracy, based on average SSIM, PSNR, MSE, and NRMSE.71 This feature makes PLayer lightweight and computation-efficient, suitable for deployment in embedded systems to translate live confocal Z-stack images into connectome maps.

In the Z-stack confocal microscopy, neighboring layers exhibit significant similarities, especially in downsampled images, despite varying fiber structures (see Video 2). Our deep learning approach effectively replaces traditional layer-by-layer Z-stack imaging. PLayer infers layer structures from adjacent layers, enriching $Z$-stacks and reducing image acquisition and processing times.

Interpolation and reconstruction of missing layers reflect the capability to predict complex topological properties of neural connectomes. PLayer visualizes neural extensions in 3D, which would otherwise appear as points in a coordinate system.3 The visualization of interpolated pixel points, as shown by the white arrow in Fig. 5(e), enables a clearer representation of neural fibers compared with reference images, preventing the dismissal of topological information. A “plug-and-play” raspberry Pi-based PLayer, capable of autonomous computation and display, showcases mobility and efficiency. This system provides practical advantages, including enhanced 3D imaging and resource savings during biological sample acquisition via sparse sampling and PLayer interpolation, along with a portable module for real-time confocal microscopy diagnostics.

Our findings suggest that when axial images of a 3D volume are missing, axial weighted interpolation can yield low-resolution estimates, which can then be enhanced using a super-resolution neural network, effectively converting the problem of improving axial resolution into a lateral super-resolution task.

As shown in Eq. (1), PLayer can reconstruct layers beyond the $0.5\text{-}\mu \mathrm{m}$ axial resolution limit of confocal microscopy, such as $0.25\text{-}\mu \mathrm{m}$-spaced layers. However, due to the physical limits of confocal imaging, this could not be experimentally validated.41 For demonstration purposes, we confined our 3D reconstruction to a $0.5\text{-}\mu \mathrm{m}$ axial resolution, although this does not reflect PLayer’s full potential. Although PLayer was trained and tested with rich neural fibrous scenarios, it can be further optimized for other types of cell data, where cells have different shapes. Based on established scaling laws,72^,73 increasing both the model size and dataset volume typically leads to improved performance. Likewise, we believe that as more data will be captured in the future, it will further enhance PLayer’s accuracy in reconstructing 3D organoid volumes, therefore contributing to scientific advancements in neuroscience. Future work will focus on expanding PLayer to various image types, such as those from electron microscopy, MRI, and diffusion tensor imaging, to enrich axial resolution and capture neural fiber structures more comprehensively.

1. Video 1. 3D reconstruction (MP4, 11.7 MB [URL: https://doi.org/10.1117/1.APN.4.3.036007.s1]).
2. Video 2. Comparison of various methods (MP4, 9.89 MB [URL: https://doi.org/10.1117/1.APN.4.3.036007.s2]).
3. Video 3. Deployment (MP4, 8.11 MB [URL: https://doi.org/10.1117/1.APN.4.3.036007.s3]).

Biographies of the authors are not available.