SpineDL: a Deep Learning-based approach for neuron and anatomical structure segmentation in immunofluorescence images of damaged spinal cords

In this study, we present SpineDL, an open-source deep learning (DL) approach for neurons and anatomical structure segmentation of the spinal cord in fluorescence images immunostained with NeuN and DAPI, within the context of murine models of spinal cord injury (SCI). SpineDL comprises two main modules: 1) SpineDL-Structure, for semantic segmentation of key spinal cord structures: gray matter, white matter, ependyma, and damaged tissue; and 2) SpineDL-Neuron, for instance-level identification of neuronal somas. To train the models, we developed the SpineDL dataset, a curated collection of 161 confocal images of mouse spinal cord, manually annotated by experts and organized into specific subsets. Both models are based on the HRNetV2-W48 architecture and were trained using state-of-the-art data augmentation and optimization techniques, implemented within the BiaPy framework, following an iterative refinement process driven by quantitative evaluation and expert feedback. Our results show that SpineDL achieves expert-level performance for both cases of structural segmentation and neuron identification. This work provides a robust, reproducible, and extensible platform for the spatial analysis of neurodegeneration following spinal cord injury, representing a step toward the automation of histopathological workflows in neuroscience.