Fast and Precise Hippocampus Segmentation Through Deep Convolutional Neural Network Ensembles and Transfer Learning.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Image Analysis

Background:

• Automatic hippocampus segmentation in 3D MRI is crucial for neurological research.
• Traditional multi-atlas registration methods have limitations in accuracy and speed.
• Deep learning offers a promising alternative for automated medical image segmentation.

Purpose of the Study:

• To develop a fully automatic, accurate, and fast deep learning-based segmentation method for the hippocampus in 3D MRI.
• To leverage Convolutional Neural Networks (CNNs) for improved segmentation performance.
• To investigate transfer learning techniques for enhancing CNN-based segmentation using multiple datasets.

Main Methods:

• A novel method utilizing an ensemble of three independent CNN models operating on orthogonal slices.
• Incorporation of distinct segmentation and error correction steps using Replace and Refine networks.
• Exploration of transfer learning to combine multiple datasets for improved segmentation quality.

Main Results:

• Achieved a mean Dice value of 0.9015 on the EADC-ADNI HarP dataset with a segmentation time of 14.8 seconds per volume.
• Demonstrated improved performance on the MICCAI dataset (mean Dice 0.8835) through transfer learning.
• The proposed CNN-based method favorably compared to existing methodologies in accuracy and speed.

Conclusions:

• The developed deep learning method provides a highly accurate and efficient solution for automatic hippocampus segmentation in 3D MRI.
• Transfer learning significantly enhances segmentation quality by effectively combining data from multiple sources.
• This approach represents a substantial advancement over traditional multi-atlas registration techniques.