LeViT-META-Unet++: Fast and Efficient Encoder with Transformer Attention Unet for Medical Image Segmentation

Medical Image segmentation (MIS) acts as a pivotal task in advancing computer-assisted medical diagnosis and treatment systems but remains challenging. The task of image segmentation has achieved significant success in recent years through the use of prominent CNN-based encoder-decoder architectures, such as U-Net. Nevertheless, convolution processes locality show their limitations when it comes to learning long-range spatial linkages and global context. With hopeful results, some research has recently tried to add transformers to the encoder-decoder components; however, due to the complex computational demands of transformers, additional improvement is needed to maximise efficiency. In order to achieve quick and highly accurate medical image segmentation, we present the LeViT-META-Unet++ model, which incorporates a LeViT transformer module into a Multi-scale high efficient transformer framework based on the Unet++ architecture. The LeViT Transformer serves as a faster encoder, balancing the computational demands of transformer blocks more effectively. Additionally, the META technique enhances the models capability to capture long-range dependencies crucial for achieving precise segmentation. The architecture employs two efficient trans-blocks at different resolutions to gather both local and global features, ensuring a more comprehensive analysis of medical images and gradually merge these features, leading to highly accurate segmentation outcomes in decoder layers. Extensive experiments on challenging medical segmentation datasets, including the MSD, BraTS and Kvasir-SEG dataset for polyp segmentation, demonstrate that LeViT-META-Unet++ exceed several SOTA methods. The suggested model achieves a superior balance between speed and accuracy, making it a promising approach for medical image segmentation tasks.