Weakly Supervised Neuron Reconstruction From Optical Microscopy Images With Morphological Priors.

Area of Science:

• Neuroscience
• Computer Vision
• Bioimaging

Background:

• Manual neuron labeling from optical microscopy (OM) images is labor-intensive and suffers from low contrast and noise.
• Limited annotated data hinders deep learning applications for neuron reconstruction.
• Existing pseudo-labeling methods introduce noise, degrading performance.

Purpose of the Study:

• To develop a deep learning approach that leverages morphological priors from existing neuron reconstructions to improve neuron extraction from OM images.
• To address the challenge of limited annotated data in neuron reconstruction tasks.

Main Methods:

• Proposed MP-NRGAN, a generative adversarial network (GAN) integrating a segmentation network with morphological prior guidance.
• The segmentation network is weakly supervised by pixel-level pseudo-labels and morphology-level supervision from reconstructed neurons.
• The discriminator network ensures extracted neuron morphology aligns with known distributions.

Main Results:

• MP-NRGAN demonstrated superior performance compared to state-of-the-art methods on the VISoR-40 and BigNeuron datasets.
• Achieved improved neuron extraction accuracy with reduced training effort.
• Effectively utilized morphological priors to guide the reconstruction process.

Conclusions:

• MP-NRGAN offers an effective solution for reconstructing neurons from challenging OM images by exploiting morphological priors.
• The method mitigates the need for extensive manual labeling, accelerating neuroscience research.
• This approach advances automated neuron reconstruction and analysis in neuroimaging.