DoGNet

DoGNet detects synapses and segments synaptic protein locations in multiplexed fluorescence imaging datasets to characterize the spatial organization and relative abundances of synaptic components.

Key Features:

• Hybrid DoG–CNN architecture: Integrates Difference of Gaussians (DoG) filters with convolutional neural networks to analyze highly multiplexed microscopy data.
• Reduced parameter count: Employs a novel architecture that results in a significantly reduced number of training parameters compared to traditional convolutional networks.
• Sample-efficient training: Trains effectively with fewer examples and reduces the risk of overfitting.
• High-throughput optimization: Architecture optimized for high-throughput investigations of synaptic features.
• Evaluated datasets: Validated on multiplexed fluorescence imaging data from primary mouse neuronal cultures and mouse cortex tissue slices.
• Performance vs conventional CNNs: Demonstrated superior performance to conventional convolutional networks when trained with a limited number of examples.
• Cross-dataset transferability: Can be efficiently transferred between datasets collected by different research groups.
• Synaptic protein segmentation: Facilitates segmentation of individual synaptic protein locations and spatial extents, revealing spatial organization and relative abundances.
• Biological scope: Applicable to analysis of presynaptic vesicles, ion channels, scaffolding and adapter proteins, and membrane receptors.

Scientific Applications:

• Synapse detection and structural characterization: Identification and classification of synapses within multiplexed fluorescence imaging datasets.
• Protein localization and spatial organization analysis: Mapping spatial extents and relative abundances of synaptic proteins at single-synapse resolution.
• High-throughput phenotypic analysis: Large-scale investigations of synaptic features across samples and conditions.
• Cross-study comparative analysis: Transfer and comparison of models and results between datasets from different research groups.

Methodology:

Integrates Difference of Gaussians (DoG) filters with convolutional neural networks in a hybrid neural network architecture, reduces the number of training parameters to enable effective training with fewer examples, outputs segmentation of synaptic protein locations and spatial extents, and was trained and evaluated on multiplexed fluorescence imaging data from primary mouse neuronal cultures and mouse cortex tissue slices.