Segmenting Brain Tissues from Chinese Visible Human Dataset by Deep-Learned Features with Stacked Autoencoder.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computer Vision

Background:

• High-resolution cryosection brain images from the Chinese Visible Human (CVH) dataset offer detailed anatomical information.
• Accurate segmentation of brain tissues (white matter, gray matter, cerebrospinal fluid) is crucial for anatomical analysis.
• Existing automated segmentation methods often fail with cryosection images due to modality differences.

Purpose of the Study:

• To develop a supervised learning-based method for segmenting brain tissues in CVH cryosection images.
• To address the limitations of current methods when applied to cryosectional data.
• To enable detailed analysis of high-resolution human brain anatomy.

Main Methods:

• Utilized a supervised learning approach employing stacked autoencoders (SAEs) for feature learning.
• Implemented a two-part model with successive three-layer SAEs to extract deep feature representations from image patches.
• Integrated a Softmax classifier for inferring tissue labels based on learned features.

Main Results:

• The proposed stacked autoencoder (SAE) method demonstrated effectiveness in segmenting CVH brain tissues.
• The SAE-based approach significantly outperformed four other classical brain tissue detection strategies.
• Successfully reconstructed three-dimensional surfaces of segmented brain tissues.

Conclusions:

• The developed supervised learning method accurately segments brain tissues in high-resolution CVH cryosection images.
• This technique offers a viable solution for analyzing anatomical structures in cryosectional brain data.
• The 3D reconstructions highlight the potential for exploring intricate human brain anatomy.