FCNN-based axon segmentation for convection-enhanced delivery optimization

Area of Science:

• Neuroscience
• Medical Imaging
• Computational Biology

Background:

• Glioblastoma multiforme treatment presents significant clinical challenges.
• Convection-enhanced delivery (CED) shows promise but is limited by drug leakage.
• Accurate brain microstructure analysis, including axon diameter distribution (ADD), is crucial for optimizing CED.

Purpose of the Study:

• To develop an automatic, accurate, and fast method for axon segmentation in electron microscopy images.
• To enable automatic computation of axon diameter distribution (ADD) for improved CED models.
• To enhance brain microstructure analysis for optimizing glioblastoma treatment strategies.

Main Methods:

• A residual fully convolutional neural network (FCNN), inspired by U-Net and Resnet, was employed for segmentation.
• Training utilized mini-batch gradient descent with the Adam optimizer.
• The Dice coefficient was selected as the loss function for model training.

Main Results:

• The FCNN achieved segmentation performance comparable to existing methods, measured by Information Theoretic Scoring.
• Training time was significantly reduced to 5 hours on a deployed GPU.
• Testing time was minimal (0.2 seconds) without requiring extensive post-processing, and computed ADDs were statistically equivalent to ground-truth data.

Conclusions:

• The developed algorithm provides a fast and accurate solution for axon segmentation and ADD computation.
• This method shows significant potential for advancing brain microstructure analysis.
• The findings support the optimization of convection-enhanced delivery (CED) for glioblastoma treatment.