A lightweight multiscale attention network for 3D tumor segmentation in PET images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate tumor segmentation in PET images is crucial for disease monitoring and treatment planning.
• Current methods struggle with insufficient spatial feature utilization in PET image segmentation.
• Developing advanced segmentation techniques is essential for improving clinical outcomes.

Purpose of the Study:

• To propose a novel multiscale attention-guided feature mechanism for 3D tumor segmentation (MA3DSeg).
• To enhance the accuracy and efficiency of tumor segmentation in PET imaging.
• To address the challenge of limited spatial feature access in PET image analysis.

Main Methods:

• Utilized a 3D-DX block for a lightweight encoder to minimize channel feature redundancy and model complexity.
• Employed a multiscale attention-guided feature enhancement (MAFE) module to fuse channel and spatial features for comprehensive tumor analysis.
• Introduced an interaction boundary semantic (IBS) module with explicit and implicit difference information to refine segmentation edges.

Main Results:

• MA3DSeg demonstrated superior performance compared to existing networks on ECPC-IDS, Hecktor 2022, and AutoPET datasets.
• Achieved a 0.5% improvement in Dice score on Hecktor 2022 and a 2.1% improvement in RVD on ECPC-IDS compared to 3D UX-Net.
• Validated excellent tumor segmentation performance and generalization ability through rigorous experiments.

Conclusions:

• MA3DSeg effectively addresses the limitations of spatial feature extraction in 3D tumor PET image segmentation.
• The proposed method offers significant improvements in segmentation accuracy and robustness.
• MA3DSeg shows great potential for clinical application in disease monitoring and treatment guidance.