Boosting Convolution With Efficient MLP-Permutation for Volumetric Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Vision Transformer (ViT) and Multi-Layer Perceptron (MLP) networks show promise in 3D volumetric medical image segmentation (Vol-MedSeg).
• Existing methods face challenges with data anisotropy and MLP resolution sensitivity.

Purpose of the Study:

• To propose PHNet, a novel permutable hybrid network for Vol-MedSeg.
• To leverage the strengths of both CNNs and MLPs for improved segmentation performance.
• To address anisotropy and resolution sensitivity in 3D medical imaging.

Main Methods:

• Developed PHNet, a hybrid network integrating 2D/3D CNNs for local feature extraction.
• Introduced a Multi-Layer Permute Perceptron (MLPP) module for long-range dependency capture and positional information preservation.
• Implemented axis decomposition and token segmentation within MLPP to handle anisotropy and resolution sensitivity.

Main Results:

• PHNet achieved state-of-the-art performance on COVID-19-20, Synapse, LiTS, and MSD BraTS benchmarks.
• The proposed method demonstrated superior results compared to existing methods.
• PHNet achieved these results with lower computational costs.

Conclusions:

• PHNet effectively combines CNNs and MLPs for robust Vol-MedSeg.
• The novel MLPP module successfully addresses key challenges in 3D medical image segmentation.
• PHNet offers a computationally efficient and high-performing solution for Vol-MedSeg tasks.