Segmentation of neuronal structures using SARSA (λ)-based boundary amendment with reinforced gradient-descent curve...

Area of Science:

• Neuroscience
• Biomedical Imaging
• Computer Vision

Background:

• Neuron structure segmentation in electron microscopy (EM) images is crucial for neurobiology.
• Low-resolution EM images present challenges like noise and limited features, hindering conventional segmentation methods.
• Accurate segmentation is vital for identifying imaging features related to brain diseases.

Purpose of the Study:

• To develop an advanced algorithm for accurate neuron structure segmentation in electron microscopy images.
• To address limitations of conventional methods in handling noisy and feature-poor EM data.
• To improve the identification of substructures for understanding brain diseases.

Main Methods:

• A multi-scale fused structure boundary detection algorithm using Gaussian pyramids and Laplacian of Gaussian (LoG) function.
• A reinforcement learning-based boundary amendment method employing SARSA(λ) for repairing gaps in detected neuron structures.
• A curve traveling and amendment approach to minimize connection costs for closing gaps.

Main Results:

• The proposed algorithm demonstrated stable and efficient structure segmentation.
• Both with and without boundary amendment, the algorithm outperformed six conventional boundary detection approaches on ISBI 2012 EM images.
• The reinforcement learning amendment significantly reduced Rand error and warping error, crucial metrics for structure segmentation.

Conclusions:

• The developed algorithm provides superior neuron structure segmentation in electron microscopy images compared to existing methods.
• The boundary amendment technique effectively addresses gaps, leading to more accurate and complete neuron structure identification.
• This method holds significant potential for advancing neurobiological research and the study of brain diseases through improved imaging analysis.