Modeling 3D Mesoscaled Neuronal Complexity through Learning-based Dynamic Morphometric Convolution

Abstract

Accurate reconstruction of neuronal morphology from three-dimensional (3D) light microscopy is fundamental to neuroscience. Nevertheless, neuronal arbors intrinsically exhibit slender, tortuous geometries with high orientation variability, posing significant challenges for standard 3D convolutions whose static, axis-aligned receptive fields lack adaptability to such complex morphology. To address this, we propose the Dynamic Morph-Aware Convolution (DMAC) framework, which incorporates inherent geometric priors into convolution by jointly adapting both the shape and orientation of the kernel. This enables morphology-aware feature extraction tailored to arborized and variably oriented neuronal trajectories. Specifically, we first apply dynamic tubular convolutions to bridge the structural mismatch between isotropic convolution kernels and the slender morphology of neurons. To sufficiently accommodate the 3D orientation variability of neuronal branches, we further introduce a rotation mechanism that dynamically reorients the tubular kernel via two learnable angles (elevation and azimuth), enabling precise alignment with local neuronal directions. We validate our method through extensive experiments on four mesoscaled neuronal imaging datasets, including two from the BigNeuron project (Drosophila and Mouse) and two additional benchmarks (NeuroFly and CWMBS). Our approach consistently outperforms state-of-the-art methods, achieving average improvements of 5.4% in Entire Structure Average (ESA), 6.9% in Different Structure Average (DSA), and 7.5% in Percentage of Different Structure (PDS). These results demonstrate the effectiveness of our proposed DMAC in capturing complex morphological variations and enhancing structural fidelity across diverse mesoscaled neuronal morphologies.