Boosting Multilabel Semantic Segmentation for Somata and Vessels in Mouse Brain.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Imaging

Background:

• Deep convolutional neural networks (DCNNs) are crucial for semantic segmentation of neural tissues in microscopy images.
• Accurate 3D reconstruction of brain vasculature and neural networks relies on high-quality human-annotated labels.
• Existing human annotations for dense nerve tissues often contain errors, limiting DCNN performance.

Purpose of the Study:

• To develop a novel boosting framework to systematically improve the quality of annotated labels for DCNN-based semantic segmentation.
• To enhance the accuracy and efficiency of 3D reconstruction of brain vasculature and neural networks.
• To address the challenge of poor-quality human annotations in neuroscience image data.

Main Methods:

• A boosting framework integrating a DCNN with a customized Dice-logarithmic loss function for multilabel semantic segmentation.
• A fusion module combining annotated labels with DCNN predictions.
• A boosting algorithm for sequential updating of sample weights during network training.

Main Results:

• The proposed framework achieved state-of-the-art performance in segmenting mouse brain somata and vessels.
• The framework demonstrated effectiveness even when trained with datasets containing poor-quality human annotations.
• Improved segmentation task performance was observed through enhanced label quality.

Conclusions:

• The developed AI technique can significantly advance neuroscience research by improving image segmentation.
• This boosting framework offers a robust solution for handling imperfect annotations in biomedical image analysis.
• The approach facilitates more accurate and efficient 3D reconstruction of complex neural structures.