An Efficient Vision Mamba-Transformer Hybrid Architecture for Abdominal Multi-Organ Image Segmentation

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Medicine
• Computer Vision

Background:

• Accurate abdominal multi-organ segmentation is crucial for clinical applications like disease diagnosis and treatment planning.
• Existing deep learning models face challenges in balancing segmentation accuracy with computational efficiency, especially with complex image data.
• Inhomogeneous intensity distributions and intricate anatomical structures pose significant hurdles for current segmentation techniques.

Purpose of the Study:

• To develop a hybrid deep learning framework for improved abdominal multi-organ segmentation.
• To enhance both segmentation accuracy and computational efficiency in medical image analysis.
• To address limitations of current methods in handling challenging image characteristics.

Main Methods:

• Integration of an Efficient Vision Mamba (EViM) module within a Transformer-based encoder architecture.
• Leveraging hidden-state mixer-based state-space duality in the EViM module for efficient global context modeling and channel-wise interactions.
• Utilizing a weighted combination of cross-entropy and Jaccard loss to refine boundary delineation.

Main Results:

• Achieved an average Dice score of 82.67% and an HD95 of 16.36 mm on the Synapse dataset, surpassing state-of-the-art methods.
• Demonstrated generalizability across different imaging modalities through validation on the ACDC cardiac MR dataset.
• Successfully integrated global and local information for robust segmentation performance.

Conclusions:

• The proposed hybrid framework offers a practical and robust solution for clinical abdominal multi-organ segmentation.
• The model achieves high segmentation accuracy and computational efficiency, addressing key challenges in the field.
• The approach shows promise for advancing automated medical image analysis and supporting clinical decision-making.