DDeep3M+: adaptive enhancement powered weakly supervised learning for neuron segmentation.

Area of Science:

• Neuroscience
• Computational Biology
• Medical Imaging

Background:

• Accurate 3D neuron segmentation is vital for understanding neural circuits and brain function.
• Challenges in optical microscopy (OM) images, such as noise and low contrast, hinder precise neuron segmentation.
• Current deep learning methods (CNNs) require extensive manual labels, limiting their application and generalization.

Purpose of the Study:

• To develop a weakly supervised learning framework for neuron segmentation that eliminates the need for manual labeling.
• To improve the accuracy and generalization of neuron segmentation algorithms for 3D OM datasets.
• To facilitate large-scale neuronal reconstruction and morphological studies.

Main Methods:

• A novel weakly supervised framework, DDeep3M+, was developed for automated neuron segmentation.
• Hessian analysis-based adaptive enhancement filters were used to generate initial pseudo-labels.
• Iterative refinement of pseudo-labels and retraining of the DDeep3M model were performed to enhance segmentation accuracy.

Main Results:

• The DDeep3M+ method achieved a high performance score of 0.973, comparable to supervised CNN models.
• The proposed method demonstrated superior performance compared to existing segmentation algorithms.
• Results indicate strong generalization capabilities on 3D OM datasets.

Conclusions:

• DDeep3M+ offers an accurate and efficient solution for neuron segmentation without manual annotation.
• The method significantly advances the potential for large-scale neuronal reconstruction and brain research.
• Weakly supervised learning presents a viable alternative for overcoming data limitations in neuroimaging analysis.