3D Capsule Networks for Brain Image Segmentation.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in radiology
• Neuroimaging

Background:

• Current autosegmentation models like UNets and nnUNets struggle with segmenting novel brain MRI data and lack computational efficiency.
• 3D capsule networks present a promising alternative to overcome these limitations in medical image segmentation.

Purpose of the Study:

• To develop and validate a 3D capsule network for brain image segmentation.
• To evaluate its performance against UNets and nnUNets in terms of accuracy, robustness to underrepresented data, and computational efficiency.

Main Methods:

• Utilized 3430 multi-institutional brain MRIs for training and validation.
• Compared a 3D capsule network against UNets and nnUNets using Dice scores and computational metrics.

Main Results:

• Capsule network achieved high Dice scores (95% for third ventricle, 94% for thalamus, 92% for hippocampus), comparable to UNets/nnUNets.
• Significantly outperformed UNets by 30% in segmenting underrepresented images.
• Required less than one-tenth the computational memory and was over 25% faster to train than UNets/nnUNets.

Conclusions:

• A validated 3D capsule network demonstrates effectiveness in brain image segmentation.
• The capsule network excels at segmenting images not well-represented in training data.
• This approach offers superior computational efficiency compared to existing models.