CLT-MambaSeg: An integrated model of Convolution, Linear Transformer and Multiscale Mamba for medical image...

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Deep learning advances medical image segmentation, but balancing feature localization, global context, and efficiency is challenging.
• Convolutional Neural Networks (CNNs) excel at local features but struggle with long-range dependencies.
• Transformers capture global context but are computationally intensive and data-hungry.

Purpose of the Study:

• To propose CLT-MambaSeg, a novel architecture for medical image segmentation.
• To integrate Convolution, Linear Transformer, and Multiscale Mamba for improved feature localization, global context modeling, and computational efficiency.
• To address limited data challenges with a Memory-Guided Augmentation Generative Adversarial Network (MeGA-GAN).

Main Methods:

• Developed CLT-MambaSeg integrating Convolution, Linear Transformer, and Multiscale Mamba.
• Employed a convolution-based Spatial Representation Extraction (SREx) module for intricate spatial relationships.
• Utilized a Mamba Vision Linear Transformer (MVLTrans) module for multiscale and global context.
• Introduced MeGA-GAN for synthetic data generation to augment limited datasets.

Main Results:

• Extensive experiments on five benchmark datasets (CVC-ClinicDB, BUSI, PH2, ISIC-2016, ISIC-2017).
• CLT-MambaSeg demonstrated superior performance compared to existing state-of-the-art methods.
• Ablation studies confirmed the efficacy of individual components and the overall architecture.

Conclusions:

• CLT-MambaSeg effectively captures local and global features while maintaining computational efficiency in medical image segmentation.
• The proposed method, enhanced by MeGA-GAN, shows significant potential for improving medical image segmentation accuracy, especially with limited data.
• CLT-MambaSeg represents a promising advancement in deep learning for medical image analysis.