Dynamic Convolution Model Enhances 3D Reconstruction of Complex Mesoscale Neurons

Researchers have developed Dynamic Morph-Aware Convolution, an advanced deep-learning framework designed to reconstruct complex three-dimensional neuronal morphology from volumetric light microscopy by dynamically adapting kernel shape and spatial orientation. Accurately digitizing slender, tortuous neural branches poses fundamental computational challenges for standard 3D convolutional neural networks, which rely on rigid, axis-aligned receptive fields that fail to accommodate filamentous geometries. To resolve this structural mismatch, the proposed architecture combines elongated tubular kernel stretching, learnable elevation and azimuth rotation mechanisms, and cumulative spatial deformation. Across extensive empirical evaluations on four mesoscale neuronal imaging benchmarks, the framework consistently outperformed established state-of-the-art segmentation and tracing baselines. The method achieved average tracing improvements of 5.4 percent in Entire Structure Average, 6.9 percent in Different Structure Average, and 7.5 percent in Percentage of Different Structure. These results demonstrate that embedding explicit geometric priors directly into convolutional feature extraction substantially preserves structural continuity and elevates topological fidelity in volumetric connectomics.

Reconstructing neuronal morphology is essential for analyzing neural circuits and understanding brain network activity. In volumetric microscopy, tracing filamentous neuronal projections historically required labor-intensive manual annotation by neuroscientists, creating a bottleneck for large-scale morphology analysis. Community initiatives like the DIADEM and BigNeuron challenges, alongside open repositories such as NeuroXiv, were established to advance automated computational tracing algorithms. However, tracing complex neuronal morphologies remains challenging because microscopic neuron images exhibit extreme thinness, high curvature, and substantial directional variability. Early automated reconstruction methods depended on handcrafted algorithms categorized into seed-point dependent approaches and seed-point independent trajectory exploration, often supported by classical filters such as Hessian matrices, support vector machines, ray-shooting models, or region growing. Although deep learning models like 3D U-Net and V-Net introduced data-driven adaptability, standard isotropic cubic kernels inherently struggle with neuronal arbors. Their static receptive fields gather excessive background noise while failing to capture filamentous, orientation-varying patterns.

To address the structural mismatch between isotropic kernels and slender neuronal morphology, the researchers restructured conventional three-dimensional convolution through a tubular stretching formulation. Standard convolutional layers deploy isotropic cubic kernels, such as three-by-three-by-three grids, distributing spatial receptive fields uniformly across dimensions. While effective for modeling blob-like anatomical structures, isotropic kernels are inherently suboptimal for slender neurites, as they incorporate excessive background information while failing to aggregate contextual signals along extended fiber paths. Inspired by prior anisotropic kernel explorations, the Dynamic Morph-Aware Convolution framework flattens and extends the standard three-by-three-by-three kernel into an elongated one-by-one-by-twenty-seven tubular receptive field aligned along a chosen principal axis. In this design, sampling points are arranged symmetrically around the central grid coordinate across discrete positions, establishing an elongated sampling window of two times the radius plus one. This stretched geometry provides natural alignment with linear neuronal segments, improving feature extraction along continuous filaments while suppressing background noise.

Because fixed-axis tubular stretching alone cannot navigate sharp bends or arbitrarily oriented neural branches, the authors introduced a dynamic rotation mechanism. Neuronal projections rarely align with Cartesian axes; instead, branches twist and extend obliquely across three-dimensional tissue volumes. When encountering abrupt trajectory shifts, a static tubular receptive field suffers from severe orientation mismatch, leading to degraded feature representations and fragmented segmentations. The Dynamic Morph-Aware Convolution architecture overcomes this constraint by predicting voxel-wise rotation maps directly from input feature maps using a 3D convolutional layer coupled with group normalization and a softsign activation function. This operation outputs two continuous learnable angles, an elevation angle and an azimuth angle, scaled by pi over two to capture vertical tilt and horizontal rotation. Multiplying elementary elevation and azimuthal rotation matrices constructs a local 3D rotation matrix at every voxel, dynamically rotating all kernel sampling coordinates around the central point to follow local neuronal orientations.

Following dynamic rotational alignment, the architecture applies a learnable cumulative deformation mechanism to capture subtle local structural variations while preserving kernel continuity. Standard deformable convolutions predict independent offsets for each sampling point, causing points to scatter across disjointed positions and detach from slender targets. To avoid kernel dispersion and preserve structural integrity, the Dynamic Morph-Aware Convolution module implements a recursive offset strategy restricted strictly to the two non-principal axes, keeping coordinates along the principal axis stationary. Incremental offsets are predicted for each sampling location and bounded within negative one to positive one. The network recursively accumulates these bounded offsets outward from the kernel center in both forward and backward directions, ensuring smooth spatial curvature that mirrors natural neurite bending without destabilizing tubular geometry. Because deformed sampling positions rarely coincide with discrete voxel grid coordinates, the framework utilizes trilinear interpolation over the feature volume, ensuring fully differentiable end-to-end gradient backpropagation throughout training.

To enrich spatial representation and overcome the limitation of analyzing volumetric tissues from a single orientation, the framework incorporates an orthogonal multi-view aggregation strategy. Volumetric neuronal structures display intricate geometric complexity that cannot be fully captured from one principal axis. The architecture therefore deploys three parallel Dynamic Morph-Aware Convolution modules, each oriented along one of the canonical Cartesian axes: x, y, and z. Each axis-specific branch performs independent tubular stretching, dynamic rotational reorientation, and cumulative deformation. The resulting directional feature representations are concatenated along the channel dimension and integrated through a fusion operator consisting of a one-by-one-by-one convolution, normalization, and activation layer. This modular architecture integrates into a U-shaped encoder-decoder network featuring MedNeXt-style blocks and cross-scale skip connections. To further stabilize optimization, the training pipeline applies deep supervision across multiple intermediate network stages, utilizing an objective function combining multi-scale Dice loss for volumetric overlap optimization and Cross-Entropy loss for precise voxel-wise binary classification.

The computational framework was validated across four diverse mesoscale neuronal reconstruction datasets representing distinct imaging modalities, signal qualities, and biological species. These benchmarks included two subsets from the BigNeuron project: a Drosophila dataset of forty-two volumetric stacks and a Mouse dataset containing twenty-two high-resolution volumes. The Drosophila test volumes averaged approximately 202 by 254 by 273 voxels, while the anisotropic mouse volumes averaged 39 by 1140 by 1140 voxels. Additionally, the authors evaluated the system on NeuroFly, consisting of 91 training and 62 testing volumes of 128-cubed voxels obtained via VISoR and fMOST microscopy. The fourth benchmark, the Complex Whole Mouse Brain Sub-Image dataset, comprised 245 volumetric blocks of 256-cubed voxels exhibiting strong background noise or weak fluorescence signals. Across all benchmarks, volumetric images were preprocessed through resampling to 1.0 mm isotropic spacing, intensity normalization to zero mean and unit variance, random 32-cubed patch extraction, and extensive geometric augmentations.

Models were trained on two NVIDIA RTX 4060 GPUs using AdamW with an initial learning rate of 0.001 across 110 epochs with automatic mixed precision. Inference was executed using a sliding-window strategy matched to the training patch size. In quantitative segmentation assessments against baselines like SegMamba, TransUNet, nnUNet, and MedNeXt, the architecture achieved top performance across all four datasets. On Drosophila, the model attained an F1 score of 49.04 percent and a Hausdorff distance of 3.02, outperforming the closest baseline by 1.47 percentage points in F1 and lowering Hausdorff distance by 0.11. On Mouse, it recorded an F1 score of 53.82 percent and a Hausdorff distance of 7.90, improving F1 by 1.77 percentage points. On NeuroFly, it reached an F1 score of 64.95 percent and a Hausdorff distance of 16.74, while on CWMBS it achieved an F1 score of 37.29 percent and a Hausdorff distance of 14.52.

The downstream tracing utility was evaluated using two automated pipelines, SmartTracing and NeuTube, measuring Entire Structure Average, Different Structure Average, and Percentage of Different Structure. Across both tracing algorithms, the framework achieved lowest error metrics, demonstrating improved structural completeness and topological alignment to ground truth. Systematic ablation experiments on the Drosophila and Mouse datasets confirmed the individual contribution of each architectural component. Comparing the full model against variants with a single rotation angle or zero rotation confirmed that full two-angle orientation adaptability is vital for tracking directionally heterogeneous branches. Crucially, replacing all Dynamic Morph-Aware Convolution blocks with standard 3D convolutions caused pronounced performance degradation; on Drosophila evaluated with SmartTracing, Entire Structure Average deteriorated from 1.42 to 1.80, Different Structure Average rose from 4.20 to 4.43, and Percentage of Different Structure escalated from 0.18 to 0.23. Further ablations showed that expanding kernel size from three-cubed to five-cubed oversmoothed delicate fibers, while disabling multi-view aggregation or deep supervision consistently impaired accuracy.

Assessing computational efficiency, researchers noted that Dynamic Morph-Aware Convolution delivers superior topological precision without excessive model scaling. Unlike MedNeXt-Large, which expands parameters and channel dimensions yet degraded on multiple datasets, the proposed architecture focuses computation on geometric alignment. Consequently, the model achieved substantially higher segmentation and reconstruction accuracy while operating with 20 percent fewer parameters than MedNeXt-Large. Although evaluating multiple deformed tubular kernels across orthogonal anatomical axes introduces a moderate increase in inference latency, the authors observed that this trade-off remains highly advantageous for connectomics, where topological fidelity and structural continuity strictly take precedence over marginal runtime differences. The authors also delineated boundary conditions, noting that while tubular geometric priors excel at resolving slender, curvilinear neuronal arbors, this morphological specialization may constrain direct applicability to non-tubular or irregular biological structures. Nonetheless, for high-throughput neuron reconstruction, integrating deformable tubular stretching, dynamic 3D rotational tracking, and multi-view aggregation establishes an effective paradigm for volumetric biomedical imaging.

Subject of Research: Deep learning-based 3D neuronal morphology reconstruction and segmentation

Article Title: Modeling 3D mesoscaled neuronal complexity through learning-based dynamic morphometric convolution

Article References: Cheng, Y. S., Zhao, R., Wang, H., Peng, H., Chrzanowski, W., & Cai, W. (2026). Modeling 3D mesoscaled neuronal complexity through learning-based dynamic morphometric convolution. Brain Informatics, 13(1), Article 4. https://doi.org/10.1186/s40708-025-00288-5

Image Credits: AI Generated

DOI: 10.1186/s40708-025-00288-5

Keywords: neuron reconstruction, neuronal morphology, light microscopy, connectomics, Dynamic Morph-Aware Convolution, 3D convolution, tubular kernel, deformable convolution, deep supervision, multi-view aggregation, BigNeuron, NeuroFly, CWMBS, SmartTracing, NeuTube, volumetric image segmentation, Hausdorff distance