3D MFA: An automated 3D Multi-Feature Attention based approach for spine segmentation using a multi-stage network...

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Spine segmentation is challenging due to complex anatomy and imaging variations.
• Unclear boundaries and tissue overlap complicate accurate segmentation.
• Existing methods struggle with diverse imaging modalities and anatomical variability.

Purpose of the Study:

• To propose a novel 3D Multi-Feature Attention (MFA) model for robust spine segmentation.
• To enhance feature representation for improved segmentation accuracy.
• To achieve state-of-the-art performance with reduced computational overhead.

Main Methods:

• A novel 3D Multi-Feature Attention (MFA) model was developed for spine segmentation.
• Standard MobileNetv3 was enhanced with Reverse Convolution Block Attention Module (RCBAM) and Feature Pyramid Pooling (FPP).
• Separate training on axial, coronal, and sagittal views, followed by feature concatenation for 3D segmentation.

Main Results:

• The 3D MFA model achieved high performance on the VerSe 2020 and VerSe 2019 datasets.
• Dice Coefficient Score (DCS) of 96.52% and Intersection over Union (IoU) of 95.84% on VerSe 2020.
• DCS of 94.64% and IoU of 93.69% on VerSe 2019, outperforming state-of-the-art methods.
• Demonstrated superior segmentation accuracy with lower computational cost.

Conclusions:

• The proposed 3D MFA model offers a significant advancement in spine segmentation.
• The model effectively handles anatomical complexity and imaging variability.
• Achieves competitive results with improved efficiency, making it suitable for clinical applications.