Cascaded Spatial and Depth Attention UNet for Hippocampus Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Accurate hippocampus segmentation is crucial for diagnosing neurological disorders.
• Existing segmentation methods face challenges with small, asymmetrical brain structures.
• Deep learning, particularly UNet architectures, shows promise in medical image analysis.

Purpose of the Study:

• To introduce and evaluate the novel Cascaded Spatial and Depth Attention U-Net (CSDA-UNet) for precise hippocampus segmentation.
• To enhance UNet by integrating spatial and inter-slice attention mechanisms for improved volumetric accuracy.
• To assess the model's performance on benchmark datasets like ADNI and Decathlon.

Main Methods:

• Developed the CSDA-UNet architecture incorporating Spatial Attention (SA) and Inter-Slice Attention (ISA) modules.
• Utilized T1-weighted brain MRI scans from the Alzheimer's Disease Neuroimaging Initiative (ADNI) and Decathlon datasets.
• Evaluated segmentation performance using Dice coefficient and Intersection over Union (IoU) metrics.

Main Results:

• CSDA-UNet achieved a Dice coefficient of 0.9512 and IoU of 0.9345 on the ADNI dataset.
• On the Decathlon dataset, CSDA-UNet obtained Dice scores of 0.9907 (train) and 0.8963 (validation).
• The model demonstrated superior performance compared to state-of-the-art methods in segmenting the hippocampus.

Conclusions:

• The proposed CSDA-UNet effectively segments the hippocampus, accurately capturing small, asymmetrical structures.
• The dual-attention framework enhances volumetric uniformity and inter-slice dependency modeling.
• CSDA-UNet offers computational efficiency suitable for clinical deployment in neuroimaging research.