Adaptive Decomposition and Shared Weight Volumetric Transformer Blocks for Efficient Patch-Free 3D Medical Image...

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• High-resolution 3D medical image segmentation is crucial for diagnosis but computationally challenging.
• Existing methods often compromise global context or require complex training.
• Patch-sampling reduces computational cost but degrades performance.

Purpose of the Study:

• To develop a cost-effective and high-performance 3D medical image segmentation framework.
• To overcome limitations of patch-based and existing patch-free methods.
• To improve segmentation accuracy while maintaining computational feasibility.

Main Methods:

• Introduced Adaptive Decomposition (A-Decomp) to reduce feature spatial size and memory consumption.
• Developed Shared Weight Volumetric Transformer Blocks (SW-VTB) for efficient long-range dependency capture.
• Proposed a novel patch-free segmentation framework, VolumeFormer, combining A-Decomp and SW-VTB.

Main Results:

• VolumeFormer demonstrated superior performance compared to existing patch-based and patch-free methods on two datasets.
• The proposed method achieved a comparatively fast inference speed.
• The framework exhibits a relatively compact design with reduced parameter numbers.

Conclusions:

• VolumeFormer offers an effective solution for high-resolution 3D medical image segmentation.
• The novel approach balances performance, computational cost, and efficiency.
• This patch-free framework advances the field of medical image analysis.