CTCF-UNet: 3D Concurrent Transformer-CNN Fusion U-Net for Brain Tumor Segmentation

Accurately segmenting brain tumors is essential for diagnostic and planning treatment. Convolutional neural networks (CNNs) have recently displayed noteworthy results in segmenting brain tumors. However, they have a poor capability to extract global features. Conversely, Transformers can effectively extract global features but are unable to adequately capture local features. To segment brain tumors accurately, it is vital to extract both global and local features. Therefore, several works have designed innovative architectures by combining CNN and Transformer to exploit their advantages. Despite the promising performances of these Transformer-based networks, it remains challenging to create a network architecture that can capture global and local features effectively. To tackle this issue, we present a new Transformer-based 3D U-Net variant termed 3D Concurrent Transformer-CNN Fusion U-Net (CTCF-UNet) that exploits the benefits of Swin Transformer and CNN to effectively utilize global and local features. CTCF-UNet enhances 3D U-Net by introducing a Concurrent Transformer Convolution Fusion (CTCF) block as the encoder and decoder blocks. The CTCF block concurrently extracts global and local features using Swin Transformer and convolutional blocks and subsequently fuses these complementary features to enhance the feature representation capability. Comprehensive experiments on the BraTS 2021 and 2020 benchmark datasets prove the success of the introduced technique. The proposed CTCF-UNet outperformed the CNN-and Transformer-based state-of-the-art networks.