MSA²-Net: Utilizing self-adaptive convolution module to extract multi-scale information in medical image...

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• nnU-Net framework has limitations in flexibility due to fixed configurations, especially convolution kernel sizes.
• 3D medical imaging presents challenges due to continuous spatial evolution of anatomical structures along the Z-axis.

Purpose of the Study:

• Introduce a self-adaptive convolution module to address limitations in existing frameworks for 3D medical image segmentation.
• Enhance the ability of deep learning models to capture dynamic structural transformations of organs.

Main Methods:

• Developed a self-adaptive convolution module using a differentiable soft-attention mechanism to aggregate candidate kernels.
• Integrated the module into the multi-scale convolution bridge and multi-scale amalgamation decoder of the MSA²-Net architecture.
• Employed the module to refine features, align with spatial continuity, and precisely reconstruct organ details.

Main Results:

• MSA²-Net achieved competitive Dice scores: 86.49% (Synapse), 92.56% (ACDC), 93.37% (Kvasir), and 92.98% (ISIC2017).
• The self-adaptive module improved the capture of global context and local nuances in feature maps.
• Demonstrated robustness in handling complex spatial variations across diverse medical imaging modalities.

Conclusions:

• The proposed self-adaptive convolution module enhances segmentation accuracy by adapting to dynamic anatomical changes.
• MSA²-Net offers a flexible and robust solution for 3D medical image segmentation, outperforming fixed-configuration models.
• The model's ability to preserve topological intricacies is crucial for precise organ detail reconstruction.