SynAPSeg: A novel dataset and image analysis framework for deep learning-based synapse detection and quantification

Synapses are the fundamental units of neural computation, yet quantifying their organization across circuit-level scales remains a critical bottleneck in neuroscience. While advances in fluorescent labeling and imaging can generate vast datasets, analysis is often the limiting factor. Several deep learning-based tools have been proposed to ameliorate these issues. However, existing applications primarily focus on dendritic spines and lack robust solutions for segmenting synaptic puncta in dense tissue preparations. To address this, we introduce SynAPSeg, which encompasses an open-source framework for deep learning-based analysis and, to the best of our knowledge, the first large-scale, publicly available instance segmentation dataset specifically curated for synaptic puncta. We use this dataset to train deep learning models that reach the performance of human experts across a unique benchmark dataset. SynAPSeg integrates these models into an interactive interface, with support for multi-dimensional data, enabling fully automated segmentation and quantification pipelines alongside an annotation module for refinement and validation. We demonstrate the framework's scalability by performing the first comprehensive mapping of nearly 4 million excitatory postsynaptic PSD95 puncta within inhibitory interneurons across the dorsal hippocampus, revealing regional differences in synapse properties. Finally, we show SynAPSeg’s utility for 3D quantification by applying these models to study aging-associated synaptic changes in CA1 parvalbumin (PV)-positive inhibitory neurons. Through this approach, we uncover a reduction in PSD95 density along PV dendrites in the aged CA1, indicating reduced glutamatergic recruitment of PV neurons which could contribute to age-related cognitive decline. Collectively, these results demonstrate that SynAPSeg provides a scalable solution for comprehensively studying synaptic architecture in health and disease.Author summary: The connections between brain cells, known as synapses, are the fundamental units of learning and memory. Synapses are often densely packed along dendrites making them difficult to accurately segment with automated image analysis methods. The current gold standard is manual synapse identification, but this is painstakingly slow and highly subjective. While artificial intelligence has shown promise in analyzing biological images, there has been a lack of specialized training data to teach these programs to identify synapses. To address this, we created SynAPSeg, an accessible software platform that includes a large-scale, publicly available collection of expert-annotated synapse images. We used this dataset to train deep learning models to accurately identify synapses in a fraction of the time of human experts–across a range of imaging conditions and sample preparations. In proof-of-principle experiments, we use SynAPSeg to map millions of synapses across the hippocampus to better understand how excitatory input onto inhibitory interneurons changes with aging. Our results indicated a significant decrease in the number of excitatory synapses within parvalbumin-positive inhibitory interneurons, which could be relevant to age-related cognitive decline. In summary, SynAPSeg is an open-source tool that provides a fast, reliable method to automatically segment synapses in fluorescent imaging datasets.