U-MLP: MLP-based ultralight refinement network for medical image segmentation.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Medicine
• Deep Learning for Healthcare

Background:

• Convolutional Neural Networks (CNNs) and Transformers are vital in medical AI but have limitations.
• CNNs struggle with long-range dependencies, while Transformers face computational complexity and parameter challenges.
• Multi-Layer Perceptron (MLP)-based networks offer a promising alternative, achieving high accuracy with reduced computational load.

Purpose of the Study:

• To introduce U-MLP, an encoder-decoder network leveraging the ReMLP block for enhanced medical image segmentation.
• To address the limitations of existing CNN and Transformer models in medical image processing.
• To improve accuracy and efficiency in computer-aided diagnosis and intelligent medicine applications.

Main Methods:

• Developed U-MLP, an encoder-decoder network utilizing the ReMLP block.
• The ReMLP block integrates an overlapping sliding window for local feature extraction and a Multi-head Gate Self-Attention (MGSA) module for multi-dimensional information fusion.
• Incorporated the Vague Region Refinement (VRRE) module to enhance model generalization by refining pixel classification based on feature proximity.

Main Results:

• U-MLP demonstrated superior performance in medical image segmentation tasks.
• Achieved Dice Similarity Coefficients of 88.27% for skin lesions, 97.61% for spleen, and 95.91% for left atrium segmentation on benchmark datasets.
• Outperformed 7 state-of-the-art methods across the evaluated segmentation tasks.

Conclusions:

• U-MLP offers a more efficient and accurate approach to medical image segmentation compared to traditional CNNs and Transformers.
• The proposed architecture effectively captures both local and contextual semantic information, leading to improved segmentation performance.
• U-MLP shows significant potential for advancing computer-aided diagnosis and intelligent medicine.