Accurate 3D cell nuclei segmentation is often the first step in the quantitative analysis of biological imaging data, especially in intricate neural systems. Despite advancements in 3D segmentation networks, existing models still struggle to capture the intricate structures of neuronal nuclei. This limitation arises from their reliance on sparse skip connections and fixed filter-based upsampling, which hinder effective multi-scale feature fusion and fine detail recovery. To address this challenge, we propose Swin Dense Dynamic Fusion network (SwinDDF), a novel architecture specifically designed for segmenting 3D nuclei with complex shapes. SwinDDF integrates a Dense Decoding Strategy (DDS) that efficiently fuses multi-scale features through dense skip connections, combined with an innovative Dynamic Feature Fusion Module (DFFM) that leverages dynamically generated filters to enhance detail recovery and suppress noise. In parallel, we present the MEC-Nuclei dataset (derived from MEC dataset), a meticulously annotated collection of volumetric electron microscopy (vEM) images focused on the nuclei of the mouse Medial Entorhinal Cortex (MEC), a brain region critical for memory and navigation. The dataset contains various types of cell nuclei with diverse shapes, making it a challenging yet valuable resource for 3D nuclei segmentation tasks. We demonstrate the effectiveness of SwinDDF through comprehensive evaluations on the MEC-Nuclei, NucMM-Z, and NucMM-M datasets. Experimental results demonstrate the effectiveness of our method. Furthermore, we perform large-scale 3D reconstruction and report geometric feature statistics of MEC nuclei, providing valuable structural insights to support neuroscience research. Code and data are available at https://github.com/xieh01/MEC_Nucleus_Seg .