Deep 3D Neural Network for Brain Structures Segmentation Using Self-Attention Modules in MRI Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Deep learning models are increasingly used in healthcare but often lose data with pooling operations.
• Existing models focus on local information, neglecting valuable global context in medical image segmentation.

Purpose of the Study:

• To propose a deep learning approach for medical image segmentation that effectively utilizes both global and local features.
• To improve the accuracy and efficiency of brain structure segmentation using advanced neural network architectures.

Main Methods:

• A novel deep learning architecture was developed, employing convolutional neural network (CNN) layers without pooling to extract local features from 3D magnetic resonance image blocks.
• Self-attention modules were integrated to capture global context from the extracted feature maps.
• The model was trained on the Mindboggle-101 dataset, utilizing an upsampling-heavy decoder pipeline.

Main Results:

• The proposed model achieved a higher Mean Dice Score of 0.90 ± 0.036 compared to other UNet-based approaches.
• The average segmentation time was significantly reduced to approximately 0.038 seconds per brain structure.
• The model successfully segmented 37 brain structures, representing the largest number segmented in a 3D approach using attention mechanisms.

Conclusions:

• The integration of self-attention modules effectively incorporates global context, leading to more precise and faster medical image segmentation.
• The developed deep learning approach offers a robust solution for brain structure segmentation, outperforming existing methods in accuracy and speed.