Glancing Beyond Patch: Spatial Contextual Cues for 3D Neuron Segmentation.

Area of Science:

• Neuroscience
• Computational Biology
• Image Analysis

Background:

• Accurate neuron segmentation in 3D fluorescence microscopy is crucial for neuroscience.
• Current patch-based methods struggle with global neuronal morphology, leading to segmentation discontinuities.
• Reconstructing neuronal structures is hindered by fragmented segmentations.

Purpose of the Study:

• To develop a novel deep learning architecture for accurate 3D neuron segmentation.
• To address the limitations of patch-based processing by incorporating global contextual information.
• To improve neuronal reconstruction by preserving global structures and fine details.

Main Methods:

• Proposed a dual U-Net architecture, Glancing Beyond Patch Network (GBP-Net).
• Integrated contextual and high-resolution information using two U-Nets.
• Employed a cross-scale context module (CSCM) with cross-attention and a cross-resolution fusion module (CRFM) with Mamba.
• Introduced a cross-network loss function to focus on challenging segmentation samples.

Main Results:

• GBP-Net outperformed advanced segmentation methods on three datasets.
• Achieved the highest F1 scores, demonstrating superior segmentation accuracy.
• Successfully preserved global neuronal structures while capturing fine-grained details.
• Maintained computational efficiency compared to existing methods.

Conclusions:

• GBP-Net effectively incorporates global context into 3D neuron segmentation.
• The proposed architecture significantly improves segmentation accuracy and neuronal reconstruction.
• GBP-Net offers a computationally efficient and accurate solution for neuron segmentation in fluorescence microscopy.