SEF-UNet: advancing abdominal multi-organ segmentation with SEFormer and depthwise cascaded upsampling

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Abdominal organ segmentation is crucial for early disease detection and diagnosis.
• Complex anatomical structures and overlapping organs present significant segmentation challenges.
• Existing methods like CNNs and Transformers struggle with precise boundary identification and real-time performance.

Purpose of the Study:

• To develop an improved abdominal organ segmentation model with high accuracy and efficiency.
• To address the limitations of current segmentation techniques in accurately identifying organ boundaries.
• To create a model suitable for real-time medical applications with reduced computational load.

Main Methods:

• Proposed SEF-UNet, a U-Net architecture integrating SEFormer (encoder) and depthwise cascaded upsampling (dCUP) (decoder).
• SEFormer utilizes Squeeze-and-Excitation modules and depthwise separable convolutions within the MetaFormer framework.
• dCUP enhances feature integration during the upsampling process.

Main Results:

• SEF-UNet significantly improved segmentation accuracy, particularly for overlapping organ edges.
• The model demonstrated a reduction in parameter count compared to existing methods.
• Achieved superior performance in capturing fine details and texture information for precise segmentation.

Conclusions:

• SEF-UNet offers enhanced abdominal organ segmentation accuracy and efficiency.
• The model's ability to precisely delineate organ boundaries makes it promising for clinical applications.
• SEF-UNet provides a potential solution for real-time, low-latency medical image analysis.