Towards accurate abdominal tumor segmentation: A 2D model with Position-Aware and Key Slice Feature Sharing

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Abdominal tumor segmentation is vital for diagnosis but challenging due to computational demands of 3D models and data imbalance.
• Existing 2D models lack inter-slice correlation, risking tumor loss in edge slices.

Purpose of the Study:

• To propose a novel Position-Aware and Key Slice Feature Sharing 2D tumor segmentation model (PAKS-Net).
• To address the limitations of existing 2D and 3D segmentation models for abdominal tumors.
• To improve segmentation accuracy while maintaining computational efficiency.

Main Methods:

• Leveraging Swin-Transformer for global feature modeling within slices.
• Introducing a Position-Aware module to capture spatial relationships between tumors and organs.
• Employing key slices to enhance segmentation accuracy, especially for edge slices.

Main Results:

• PAKS-Net demonstrated superior performance on multiple CT datasets (KiTS19, LiTS17, pancreas, LOTUS).
• Achieved high Dice Similarity Coefficient (DSC) scores, outperforming 3D segmentation models.
• Maintained computational efficiency with fewer parameters compared to 3D models.

Conclusions:

• PAKS-Net offers an effective and computationally efficient solution for abdominal tumor segmentation.
• The model's novel approach improves accuracy by considering spatial context and inter-slice correlations.
• PAKS-Net shows promise for clinical applications in tumor screening and diagnosis.