CAST: A multi-scale convolutional neural network based automated hippocampal subfield segmentation toolbox.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Medical Image Analysis

Background:

• Accurate segmentation of hippocampal subfields is crucial for understanding brain structure and function.
• Current automated methods and manual segmentation have limitations in speed, efficiency, and reliability, particularly for smaller subfields.

Purpose of the Study:

• To develop and validate a novel automated toolbox, CAST, for accurate and efficient segmentation of hippocampal subfields.
• To assess the performance of CAST compared to existing automated methods and manual segmentation in terms of accuracy and reliability.

Main Methods:

• Development of a 3D multi-scale deep convolutional neural network (CNN) incorporating residual learning.
• Training and application of the CAST toolbox on both 7T (T2-only) and 3T (T1/T2) MRI datasets.
• Evaluation using Dice Similarity Coefficient (DSC) for accuracy and Intraclass Correlation Coefficient (ICC) for reliability.

Main Results:

• CAST demonstrated comparable segmentation accuracy (DSC) to the state-of-the-art ASHS method.
• CAST significantly improved segmentation reliability (ICC) for small subfields like CA2, CA3, and the entorhinal cortex (ERC).
• CAST achieved high consistency between DSC and ICC values across all subfields, unlike other methods.

Conclusions:

• CAST is a time-efficient, flexible, and reliable automated tool for hippocampal subfield segmentation.
• The toolbox offers improved reliability for small subfields, addressing a key limitation of existing methods.
• Publicly available code and models facilitate broader adoption and further research in neuroscience.