Attention-gated 3D CapsNet for robust hippocampal segmentation.

Area of Science:

• Neuroimaging
• Deep Learning
• Medical Image Analysis

Background:

• The hippocampus and its subfields (HSF) are crucial for memory and implicated in various neurological disorders.
• Accurate segmentation of HSF is challenging due to their small size and anatomical variability, limiting manual data labeling.
• Existing deep learning methods, like capsule networks, show promise in medical imaging but are often limited to 2D and unassessed for HSF segmentation.

Purpose of the Study:

• To introduce and evaluate a novel 3D Capsule Network (3D-AGSCaps) for automated hippocampus subfield segmentation.
• To compare the performance of 3D-AGSCaps against classical convolutional neural networks (CNNs) on segmenting HSF, particularly in small and atypical datasets.

Main Methods:

• Developed and released a public 3D Capsule Network (3D-AGSCaps).
• Compared 3D-AGSCaps to equivalent CNN architectures on three datasets with manually labeled hippocampi.
• Evaluated segmentation performance using metrics like the Dice Coefficient, especially on datasets with incomplete hippocampal inversion (IHI).

Main Results:

• 3D-AGSCaps demonstrated superior segmentation accuracy (Dice Coefficient) compared to CNNs on rotated hippocampi.
• On typical subjects, 3D-AGSCaps achieved similar Dice coefficients to CNNs but with 15 times fewer parameters (2.285M vs. 35.069M).
• These findings suggest 3D-AGSCaps can facilitate the study of atypical subjects, including those with IHI.

Conclusions:

• The developed 3D-AGSCaps offers a more accurate and automated solution for hippocampus subfield segmentation.
• This method is expected to benefit research on atypical populations, small datasets, and large cohorts where manual segmentation is impractical.
• 3D-AGSCaps holds potential for advancing the study of anatomo-functional hypotheses in conditions affecting the hippocampus.