RMTF-Net: Residual Mix Transformer Fusion Net for 2D Brain Tumor Segmentation.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Medicine
• Neuro-oncology Imaging

Background:

• Glioma surface heterogeneity and complex imaging present significant challenges for accurate medical image segmentation.
• Existing convolutional neural network methods often overlook the crucial correlation between local and global features.

Purpose of the Study:

• To develop an advanced deep learning model for precise brain tumor segmentation.
• To address limitations in current segmentation techniques by integrating local and global feature information.

Main Methods:

• Proposed the Residual Mix Transformer Fusion Net (RMTF-Net) incorporating a novel residual mix transformer encoder.
• Utilized an overlapping patch embedding mechanism within the mix transformer to preserve boundary information.
• Implemented a parallel fusion strategy with Residual Convolutional Neural Networks (RCNN) for balanced local-global feature extraction and a Global Feature Integration (GFI) module in the decoder.

Main Results:

• The RMTF-Net demonstrated superior performance compared to state-of-the-art methods on LGG, BraTS2019, and BraTS2020 datasets.
• Achieved enhanced subjective visual quality and objective evaluation metrics in brain tumor segmentation tasks.

Conclusions:

• The proposed RMTF-Net effectively segments human gliomas by fusing local and global contextual information.
• This model offers a significant advancement in medical image analysis for brain tumor segmentation.